Research proposals

PhD selection is performed twice a year through an open process. Further information are available at http://dottorato.polito.it/en/admission.

Grants from Politecnico di Torino and other bodies are available.

In the following, we report the list of topics for new PhD students in Computer and Control Engineering. This list will be regularly updated when new proposals will be available.

If you are interested in one of the topics, please contact the related Proposer.


- An index to evaluate motor fluctuations: a clinical and instrumental analysis
- Virtual and Augmented Reality: new frontiers in education and training
- Augmented Cobots: Augmented Reality to Support Collaborative Robotics
- Human-centric Advanced Driver Assistance Systems
- Big Spatio-Temporal Data Analytics
- Biological Systems Modeling for Precision Medicine
- Accessibility of Internet of Things Devices and Systems
- Security of "smart" environments
- Digital Wellbeing in the Internet of Things
- Evaluating the impact of automated decision systems in urban contexts
- Deep Natural Language Processing in Cross-lingual Domains
- From customer's devices to the cloud: orchestrating the computing continuum
- Planning Safety in the Era of Approximate Computing Systems
- Algorithms and technologies for ubiquitous applications
- Pervasive Information Management
- Key management techniques in Wireless Sensor Networks
- ICT for Urban Sustainability
- Self-evolving data-driven models
- Promoting Diversity in Evolutionary Algorithms
- Computational Intelligence for Computer Aided Design
- Visual Object Detection across Domains
- Context and Emotion Aware Embodied Conversational Agents
- Virtual character animation from video sequences
- Robust machine learning models for high dimensional data interpretation
- Urban intelligence
- Energy-Efficient Conditional Inference of Sequential Data
- Development of Agent Based Models for smart energy policies in energy communitie...
- Anomaly detection and knowledge extraction in household energy consumption
- Deep-Learning on the Edge of the IoT: optimization methods, tools and circuits f...
- Development of infrastructures, simulation models and data analytics for manufac...
- Advancing Mobile UI Testing Techniques
- Testing and Reliability of Automotive oriented System-on-Chip
- Data-powered Truck to the Future
- Convex relaxation based methods for gray-box model identification
- Stimulating the Senses in Virtual Environments
- Cross-Domain 3D Visual Learning
- Deep Learning of Object Parts and their Affordances for Task-Oriented Robot Gras...
- Cybersecurity Certifications for Hardware components and systems
- Data mining for Digital Information Literacy
- Synthesis of Smart and Intelligent Sensors
- Implementation of Machine Learning Algorithms on Ultra-Low-Power Hardware for In...
- Cybersecurity applied to embedded control systems
- Speaker verification and multi-modal identity recognition
- Quantum Computing: analysis and design of quantum algorithms for engineering app...
- Improving Quality and Reliability of Electronic Systems via accurate Defect Ana...
- Formal and non-formal learning in virtual environments
- Design of Secure Computer Hardware Architectures
- Security Enhancements in Embedded Operating Systems
- Self-supervised cross-domain activity classification from multiple information c...
- Unsupervised cross domain detection and retrieval from scarce data for monitorin...
- Advanced Programming Techniques for Massively-Parallel and Embedded Processors
- Attention-guide cross domain visual geo-localization
- Analysis, search, development and reuse of smart contracts
- Reliability and Safety of Artificial Neural Networks
- Autonomous systems' reliability and safety
- Orchestrating NFV edge services in a cloud-native world
- eXplainable Artificial Intelligence techniques for Natural Language Processing t...
- Grounded Language Learning for Multimodal Understanding
- Semantic image segmentation in open set scenarios
- Cross-domain federated graph learning
- Crowd Monitoring in the Smart City


Detailed descriptions

Title: An index to evaluate motor fluctuations: a clinical and instrumental analysis
Proposer: Gabriella Olmo
Group website: http://www.dauin.polito.it/research/research_groups/sbg_system_bi...
Summary of the proposal: This proposal is part of a large, multi-disciplinary (possibly multi-centric) cross-sectional study, to be carried out at the "Centro Parkinson e Disturbi del Movimento", Molinette" University Hospital, and in cooperation with the Dept. of Neuroscience of Università di Torino.
It is expected to explore the use of inertial sensors to monitor the circadian motor fluctuations of Parkinson's disease (PD) patients, and to answer the following questions:
-Is a wearable inertial sensor able to gather reliable information on the patient's motor fluctuations and put this in relationship with the therapeutic conditions?
- Is it possible to represent motor aspects related to the wearing-off phenomenon, as well as those due the L-dopa side effects (dyskinesia)?

The primary outcome of the study is a clinical and instrumental identification of daily motor fluctuations, and their correlation with therapeutic conditions.
Secondary outcome is to identify markers of disease progression/disability by means of inertial data.
Rsearch objectives and methods: Parkinson's disease (PD) is the second most common neurodegenerative disease worldwide, with over 10 million people affected in the world.

At present, no drug is universally recognized to stop the disease progression, and only symptomatic therapies are available. Levodopa (L-dopa) is the gold standard drug for the control of PD motor symptoms, even though, after several years of treatment, it can bring about serious side effects such as involuntary movements (dyskinesia) as well as motor response fluctuations (wearing-off, ON-OFF periods). Actually, 70% of patients develop motor fluctuations after 9 years of L-dopa therapy, and these side effects frequently occur rather early in the disease course.

The clinical approach to motor fluctuations implies careful adaptation of the drug posology to the response of each single patient, and such a personalized process is enabled by getting precise information on the patient's functioning during daily living. This requires a thorough, timely assessment of the patient's motor conditions. On the other hand, motor fluctuations are very difficult to appreciate in a medical office, and their follow-up is currently limited to poorly reliable self-reports. Hence, remote monitoring of these signs, to be implemented using low-cost instrumentation and self-managed by the patients themselves at home during daily activities, could represent a very important tool for the clinicians.

We are implementing a cross-sectional study, to be carried out at the "Centro Parkinson e Disturbi del Movimento" at the "Molinette" Hospital (University Hospital, linked to the Dept. of neuroscience of Unito), to answer the following questions:
-Is a wearable inertial sensor able to gather reliable information on the patient's motor fluctuations and put this in relationship with the therapeutic conditions?
- Is it possible to represent motor aspects related to the wearing-off phenomenon, as well as those due the L-dopa side effects (dyskinesia)?
The primary outcome of the study is a clinical and instrumental identification of daily motor fluctuations; and their correlation with therapeutic conditions. Secondary outcomes may be:
-the identification of markers of disease progression/disability by means of inertial data;
-the identification of markers related to early development of dyskinesia (so that to reduce the L-dopa dosage in early stages of the disease).
-the identification of subtypes/atypical PD cases.
Outline of work plan: We plan to enroll one hundred patients according to the following inclusion/exclusion criteria:

Inclusion criteria:
- diagnosis of idiopathic PD according to the International Movement Disorders Society diagnostic criteria
- age lower than 80 years
- presence of motor fluctuations in the last month
- Ability to walk independently
- Hoehn and Yahr score 2 or 3

Exclusion criteria:
- diagnosis of dementia

Each enrolled patient is followed in inpatients environment for a period of 6 hours during the same day. During this time, he/she is evaluated 9 different times following a standardized protocol, while walking under controlled therapeutic conditions (Off condition, scheduled times after the administration of 300 mg of dispersible formulation of levodopa/benserazide).

At each time-point, the clinical evaluation of the relevant UPDRS-Part III items is performed. During the clinical evaluation, the patient is asked to worn the inertial sensor(s), and inertial data are collected.
Then, the patient is asked to walk along a linear 10 meters path forwards, backwards, forwards forth 4 times:
1) in normal conditions;
2) with a narrowed route in the middle of the walking path;
3) performing a motor dual task (walking and taking a cup);
4) performing a cognitive-motor dual task (e.g. answering simple questions).

Moreover, the patient is asked to perform four 360⁰ turns, two for either direction.

The PhD student will cooperate in the following technical tasks:
1. Choice of the number, type and location of the Inertial sensor(s) to be worn by the patient.
2. Support in the measure implementation. To this end, the PhD student will be allowed to access the Hospital, thanks to the mobility agreement between Polito/DAUIN and the Neuroscience dept. of Unito).
3. Data analysis by means of proper Machine learning/Deep Learning algorithms.
Expected target publications: Q1 journals relevant to the Computer Science and/or Biomedical Engineering domain.
Required skills and competences:
Current funded projects of the proposer related to the proposal: No funded project.
Possibly involved industries/companies:STMicroelectronics Astel Electronic Engineering (http://www.astel.it)

Title: Virtual and Augmented Reality: new frontiers in education and training
Proposer: Andrea Sanna
Group website: http://grains.polito.it/
Summary of the proposal: In recent years there has been huge growth in investment in Augmented Reality (AR) and Virtual Reality (VR) technologies. In 2015-16, investment grew 300% to a whopping $2.3bn. Education and training will be one of the strongest use cases for AR/VR as these technologies can help us to visualize concepts in a more interactive way. When teachers struggle to communicate complex ideas and ensure people are interpreting them correctly, we can transport students from looking at something complex to being virtually immersed in a real life example. Secondly, AR/VR eliminates the need for physical materials, which can be an expensive barrier to access. Third, AR/VR simulations can provide a unique perspective about (physical) phenomena, thus enhancing learning experiences.
This proposal aims to investigate new teaching paradigms based on AR/VR. The advent of consumer devices, such as, Microsoft Hololens, Oculus Rift, and many others, opens new challenges. Students can now take advantage of technologies and tools that can strongly improve traditional teaching methodologies. On the other hand, the design and the development of AR/VR applications is still a task for software developers and teachers are not usually able to create engaging AR/VR contents.
Rsearch objectives and methods: Main goals of this Ph.D. are the design, the implementation and the validation of new AR/VR based tools able to change and improve the traditional teaching paradigms. AR and VR technologies can help students to get a more realistic 3D space comprehension, thus enhancing learning in several disciplines such as geometry, physics, chemistry, mechanics, architecture, medicine and so on.
Several issues have to be addressed in order to design and implement engaging and effective AR/VR applications for teaching purposes. First of all, the content creation has to be addressed; therefore, a framework to develop AR/VR tools without any particular coding skill has to be developed. This framework has to enable teachers to create their own applications; moreover, teachers have to be able to develop their own contents according to a “narrative”. The definition of effective storytelling plays a key role in order to design engaging AR/VR applications according to edutainment strategies: concepts of “reward”, “gratification”, “levels of difficulty”, etc. have to be introduced to involve students in a virtuous learning mechanism.
Secondly, a mechanism to maximize (and if it is possible for measuring) the transfer of knowledge has to be introduced. This is one of the main problems of existing AR/VR applications designed for education. The transfer of knowledge does depend on the user experience (UX), which can be maximized only by user-centered design techniques. The design of the interface and of the interaction paradigm are just two issues to be investigated in order to maximized the UX.
Outline of work plan: The work plan will be defined year by year; for the first year of activity, it is expected the candidate will address the following points:
- Analysis of the state-of-the-art (in this phase, both AR/VR technologies and usual teaching methodologies for a large spectrum of disciplines will be investigated). The candidate could complete this analysis step by attending to specific Ph.D. courses related to teaching methodologies. Moreover, edutainment strategies have to be also considered as they can help the design of more attractive and engaging AR/VR applications.
- Objective definition (after the analysis step, the candidate has to be able to identify challenging open problems, thus defining concrete objectives for the next steps).
- Methods (concurrently to the objective definition step, the research methodology has to be also chosen; a right mix between theoretical formulation and experimental tests has to be identified).
- Expected outcomes (research papers, scientific contributions, significance, potential applications).

During the second year, the research activity will be mainly focused on the design and implementation. In particular, the candidate will address the following points:
- Design of AR/VR tools for supporting bachelor and master students of computer animation and 3D modeling courses
- Implementation of these new teaching paradigms based on augmented and virtual reality technologies
- Expected outcomes (research papers, scientific contributions).

During the last year, the research activity will be mainly focused on tests and result analysis. In particular, new tools will be used in the “3D Modeling Course” of the Design degree and master in the “Computer Animation” course of the Computer Science degree. Results obtained in terms of number of students that will pass the exam, average final marks, and students’ opinions will be compared with data collected for previous academic years, thus obtaining a clear assessment of the impact of AR/VR technologies on learning experiences. Expected outcomes are research papers and scientific contributions.
Expected target publications: IEEE Transactions on Education
ACM Transactions on Computing Education
International Journal of Educational Research (Elsevier)
IEEE Transactions on Visualization and Computer Graphics
ACM Transaction Human Computer Interaction
Required skills and competences:
Current funded projects of the proposer related to the proposal: Not yet but a proposal named "Digital Learning" is currently under evaluation for the regional PRISM-E call.
Possibly involved industries/companies:IS_LM SrL, Archibuzz Srl, Synesthesia Srl, Conversa Srl

Title: Augmented Cobots: Augmented Reality to Support Collaborative Robotics
Proposer: Andrea Sanna
Group website: http://grains.polito.it/
Summary of the proposal: Augmented reality (AR) superimposes data or computer-generated graphics over real-world images and uses sensors and cameras to capture the operator’s motions for feedback and control. AR has been used in a large spectrum of applications: tourism, medicine, architecture, cultural heritage, military, entertainment industry and many others. Recent technological advances opened new scenarios also in (collaborative) robotics, and this might soon have a huge impact on manufacturing and logistics automation, and eventually even home and service robots.
Robot programming was traditionally done by writing code, which was time-consuming and expensive. Moreover, only people with coding skills can program robots. On the other hand, AR can completely change the paradigm, thus letting even untrained operators to interact with collaborative robots (cobots).
AR allows a human operator to understand and visualize cobots intentions; moreover, the operator can use AR to control the robot using natural, smooth movements, giving the robot precise instructions simply by doing the tasks the robot has to emulate. This new approach is ideal for cobots, which allow human operators to work directly with the robot arm without the interference of safety cages or fencing.
Rsearch objectives and methods: Main goal of this Ph.D. is the study, the design and the implementation of new human-robot interaction paradigms based on augmented reality technologies. AR is one of the pillars of Industry 4.0 and it is expected AR will change the way technicians interact with an industrial equipment. Also the interaction human-robot will deeply change; in particular, when operators and robots share the same work environment (collaborative robotics).
First of all, AR can be used to design new interfaces that allow the user to understand robot intentions: trajectories, faults, workspaces, and so on can be visualized, thus allowing technicians to trust in cobots activities. Moreover, multimodal input modes (ranging from gestures to speech) can provide users new ways to control cobots, thus replacing the traditional programming paradigms.
This Ph.D. aims to investigate these new and intriguing challenges. All activities will be performed in collaboration with COMAU a large company leader in robotics.
Outline of work plan: The work plan will be defined year by year; for the first year of activity, it is expected the candidate will address the following points:
- Analysis of the state-of-the-art of sw and hw AR technologies; moreover, multimodal interfaces will be also considered. The candidate could complete this analysis step by attending to specific Ph.D. courses related to HCI and computer vision.
- Objective definition (after the analysis step, the candidate has to be able to identify challenging open problems, thus defining concrete objectives for the next steps). Activities such as: maintenance, assembly, cobot programming and all other tasks that can take advantage from AR will be considered.
- Methods (concurrently to the objective definition step, the research methodology has to be also chosen; a right mix between theoretical formulation and experimental tests has to be identified).
- Expected outcomes (research papers, scientific contributions, significance, potential applications).

During the second year, the research activity will be mainly focused on the design and implementation of AR interfaces. In particular, the candidate will address the following points:
- Design of AR solutions for displaying robot intentions: trajectories, workspaces, possible faults, control points, and so on.
- Design of multimodal interfaces to interact with cobots (give cobots inputs without programming).
- Implementation of these new interactions paradigms based on augmented reality technologies.
- Expected outcomes (research papers, scientific contributions).

During the last year, the research activity will be mainly focused on tests and result analysis. In particular, implemented interfaces will be applied to existing cobots such as e.DO or other commercial collaborative manipulators. Moreover, also service robots will be considered in order to extend the range of application also to other categories. The aim of this phase is to gather subjective and objective data in order to compare standard interfaces with the augmented ones in terms of usability, intuitiveness, robustness and all the other KPIs that characterize a human-robot interface. Expected outcomes are research papers and scientific contributions.
Expected target publications: IEEE Transactions on Visualization and Computer Graphics
ACM Transactions Human Computer-Human Interaction
ACM Transactions on Human-Robot Interactions
International Journal of Human - Computer Studies
International Journal of Industrial Ergonomics Computer and Graphics
Required skills and competences:
Current funded projects of the proposer related to the proposal: HuManS
Possibly involved industries/companies:COMAU

Title: Human-centric Advanced Driver Assistance Systems
Proposer: Massimo Violante
Group website: http://www.cad.polito.it
Summary of the proposal: The goal of this project is human-centric advanced driver assistance systems. It focuses on the identification of techniques to monitor the behavior and the health conditions of individuals operating on vehicles to promptly activate actions (spanning from the generation of alert to the execution of autonomous driving functions, such as limiting the speed or parking) intended for guaranteeing their safety. In particular we will focus on studying the hearth rate variability to predict critical conditions such as drowsiness and the emotional state in response to different driving scenario. Two main approaches will be considered: one based on wearable devices (e.g., smart swatches), and one based on radar technologies to collect physiological parameters useful to study the heart rate variability. Signal processing techniques will then be employed for identifying the features of interest and to predict relevant behaviors (e.g., drowsiness). Finally, embedded devices will be prototyped for allowing experimental evaluation of the developed techniques. The proposal will leverage on an existing collaboration between the proponent, Sleep Advice Technologies (a recently founded start-up company), On-Semiconductior, and a group of sleep medicine medical doctors operating with Ospedale Regina Margherita di Torino and J-Medical center.
Rsearch objectives and methods: Drowsiness is one of the main causes of car accidents worldwide: 35%–45% of road accidents are caused by drowsy driving. Identification of driving drowsiness have used the following measures:
- Vehicle-based measures: deviations from lane position, movement of the steering wheel, pressure on the acceleration pedal, etc. are constantly monitored and any change in these that crosses a specified threshold indicates a significantly increased probability that the driver is drowsy
- Driver behavior measures including yawning, eye closure, eye blinking, head pose, etc., is monitored through a camera and the driver is alerted if any of these drowsiness symptoms are detected
- Physiological measures. Many studies and researches have been conducted to determine the relationship between driver drowsiness and some physiological data studied through the analysis of the respective signals: electrocardiogram (ECG) for heart rate variability, electroencephalogram (EEG) for electrical cerebral activity, electromyogram (EMG) for muscular activity, electrooculogram (EoG) for ocular movements.
The reliability and accuracy in driver drowsiness detection by using all these measures is to be considered insufficient because of some fundamental limitations:
- Vehicle-based and driver behavioral measures work in very limited conditions because they are too dependent on external factors like geometric characteristics of the road, road marking, climatic and lighting conditions. Moreover, they are susceptible to visual barriers such as driver face position and glasses wearing. Additionally, the color of the skin and the presence of the beard could influence the reconstruction of the facial contours and characteristics, including the position/movement of the eyes and the mouth.
- Physiological measures currently studied present the issue of the intrusive nature of most of the sensors they used. Moreover, many studies have determined that all these measures are poor predictors because they become clearly effective only after the driver starts to sleep, which is too late for a preventive action.
The objective of the research is to develop non-intrusive techniques to monitor the behavior of drivers and to predict the drowsiness condition in a reliable manner before drowsiness takes actually place. As monitor techniques we intend to start from an approach which measure the photo-plethysmograph (PPG) through a smart-swatch. By applying frequency-based analysis techniques, the PPG can be studied and relevant information about the activity of the sympathetic and para-sympathetic systems can be obtained and used for drowsiness prediction (as some preliminary results already suggest). We then intend to move towards a novel monitoring system based on an innovative radar technology developed by On-Semiconductor, which allows measuring the heart rate contactless. Machine learning techniques will be used to correlate information coming from the radar with those coming from PPG analysis to identify novel drowsiness prediction techniques.
Outline of work plan: Year 1
Consolidation of an already-existing PPG-based drowsiness prediction technique developed during a master thesis. The work requires algorithm development, as well as experimentation and data collection. Measures collected using a smart-watch properly instrumented with the drowsiness prediction software will be correlated with measurements collected using a medical device provided by J-Medical and analyzed with the collaboration of Ospedale Regina Margherita and SAT.
Year 2
Development of correlation techniques between PPG-based analysis, On-Semiconductor radar and medical device data. The data collected by radar will be processed and correlated with those coming from PPG-based analysis and confirmed using observation gained via a medical device. This activities will be particularly innovative as although radar technique is known to provide accurate estimation of the heart rate, it is not clear if its spectrum contains the same level of information already available in the PPG spectrum.
Year 3
Once the mathematical foundation of the radar-based drowsiness prediction is established, a prototypical set-up will be implemented, and testing will be performed. For this experimentation access to a dynamic driving simulator will be important to reproduce safely driving scenario inducing drowsiness. Through the experimentation we will evaluate the effectiveness of the prediction technique in realistic condition.
Expected target publications: IEEE/ACM Transactions
Required skills and competences:
Current funded projects of the proposer related to the proposal: Contratto di prestazione di servizi con società Sleep Advice Technologies Srl
Possibly involved industries/companies:Sleep Advice Technologies Srl, J-Medical

Title: Big Spatio-Temporal Data Analytics
Proposer: Paolo Garza
Group website: http://dbdmg.polito.it/
Summary of the proposal: The amount of spatio-temporal data rapidly increased in the last years. The big amount of collected spatio-temporal data ranges from satellite images to ground-based sensor measurements. That big amount of heterogeneous data can be profitably exploited in several domains, such as natural hazard prevention, pollution analysis, traffic planning, etc.

We are currently overloaded by spatio-temporal data. For instance, the Copernicus Earth observation programme, supported by the European Commission, collects more than 12 terabytes per day.
To transform this overload of data into valuable knowledge, we need to (i) integrate several sources, (ii) design data tailoring techniques to select relevant data related to the target analysis, and (iii) design data mining algorithms to extract knowledge offline or in near-real time. Currently, the analyses are mainly focused on one single type of data/source at a time (e.g., satellite images or ground-based measurements). The integration of several sources into big data analytics systems capable of building accurate predictive and descriptive models will provide effective support in several application domains.

The PhD candidate will design, implement and evaluate novel big spatio-temporal data analytics solutions.

An ongoing collaboration with LINKS foundation will allow the candidate to validate his/her algorithms on real use cases.
Rsearch objectives and methods: The main objective of the research activity will be the design of big data analytics algorithms and systems for the analysis of heterogeneous big spatio-temporal data (e.g., satellite images, sensor measurements), aiming at generating predictive and descriptive models.

The main issues that will be addressed are the followings.

Scalability. The amounts of big spatio-temporal data are significantly increased in the last years and some of them are singularly large (e.g., the satellite images). Hence, big data solutions must be exploited to analyze them, in particular when historical data analyses are performed.

Heterogeneity. Several heterogonous sources are available. Each source represents a different facet of the analyzed events and provides an important insight about them. The efficient integration of the available spatio-temporal data sources is an important issue that must be addressed in order to build more accurate predictive and descriptive models.

Near-real time constraint. In several domains, timely responses are needed. For example, to effectively tackle natural hazards and extreme weather events, timely responses are needed to plan emergency activities. Moreover, large amounts of streaming and time series spatial data are generated (e.g., environmental measurements) and their integration and analysis is extremely useful. The current big data streaming systems (e.g., Spark and Strom) provide limited support for real-time and incremental data mining and machine learning algorithms. Hence, novel algorithms must be designed and implemented.
Outline of work plan: The work plan for the three years is organized as follows.

1st year. Analysis of the state-of-the-art algorithms and data analytics frameworks for big spatio-temporal data. Based on the analysis of the state-of-the-art, pros and cons of the current solutions will be identified and preliminary algorithms will be designed to optimize and improve the available approaches. During the first year, descriptive algorithms, based on offline historical data analyses, will be initially designed and validated on real data related to specific domains (e.g., natural hazards) to understand how to extract fruitful patterns for performing medium- and long-term analyses/predictions.

2nd year. The design of incremental and real-time predictive models will be addressed during the second year. For instance, classification algorithms will be designed to automatically classify streaming data in real-time. They will be initially evaluated in the natural hazard management domain.

3rd year. The algorithms designed during the first two years will be improved and generalized in order to be effectively applied in different domains.

During the second/third year, the candidate will have the opportunity to spend a period of time abroad in a leading university or research center.
Expected target publications: Any of the following journals:
- ACM Transactions on Spatial Algorithms and Systems (TSAS)
- IEEE Transactions on Knowledge and Data Engineering (TKDE)
- IEEE Transactions on Big Data (TBD)
- Journal of Big Data
- Big Data Research
- IEEE Transactions on Emerging Topics in Computing (TETC)
- ACM Transactions on Knowledge Discovery in Data (TKDD)
IEEE/ACM International Conferences
Required skills and competences:
Current funded projects of the proposer related to the proposal: I-REACT - "Improving Resilience to Emergencies through Advanced Cyber Technologies", H2020 European project
Possibly involved industries/companies:Fondazione LINKS

Title: Biological Systems Modeling for Precision Medicine
Proposer: Benso Alfredo
Group website: http://www.sysbio.polito.it
Summary of the proposal: In last decade, the sudden availability of huge amount of biological data is revolutionizing the approach to understand, diagnose and cure diseases, especially those caused by misbehavior of the genetic code. The amount of publicly available data sources has grown accordingly. Unfortunately, when researchers need to extract and integrate data from different repositories in order to create usable models, a number of issues arise and the overall perception is that while each database is a very specialized source of data, it is extremely difficult to make sense of cross-database information.
This research proposal aims to investigate the design and implementation of algorithms and tools to support the study of personalized diagnosis and cure of genetic-related diseases (cancer, autism, neurological disorders, …).
Rsearch objectives and methods: The main objective of this proposal is the implementation of a modeling framework for the analysis and simulation of the regulatory interactions at the genomic level that are responsible for a number of genetic diseases. In particular, the framework will have to include:
1. a modeling language able to support the peculiar characteristics of complex biological systems (localization, encapsulation, stochasticity, hierarchy, …)
2. a formalism allowing to represent and simulate the system under study under different initial and environmental conditions.
3. A graph-analysis toolset to study the topological properties of complex biological networks
4. a simulator, enabling the study, from a quantitative point of view, of the systems dynamics
Outline of work plan: The workplan is divided in three main WPs:
1) Modeling Language: the PhD student will have to finalize the definition of a modeling language targeted on biological systems. Our group recently developed a prototype of such a language, but several additional steps need to be completed before it can be considered usable from the Life Science community. This step will require knowledge of tools and techniques for languages definition.
2) Formalism: the second WP requires the translation of the system description into a formal model. The target formalism is a particular version of Petri Nets. In this step it will be necessary to interact with Life Science researchers in order to define and implement an initial library of building blocks able to model the different biochemical mechanisms involved in the systems under study. Moreover, it will be necessary to implement an automatic tool to translate the initial description into the formal model.
3) Graphs analysis: this WP requires the implementation of a toolset able to analyze, from a topological point of view, biological networks. The type of analysis will be focused on those topological properties that have a solid biological correlation.
4) Simulator: the last WP will be focused on the development of a custom simulator able to allow the study of the systems dynamics, starting from the formal description of the system. Several types of simulations will be necessary in order to automatically explore different possible genetic defects and forecast their consequences in the cell functionalities.
Expected target publications: PlosOne
IEEE/ACM Transactions on Computational Biology and Bioinformatics
Oxford Bioinformatics
BMC Systems Biology
Required skills and competences:
Current funded projects of the proposer related to the proposal: None at the moment. A COST project is under submission
Possibly involved industries/companies:None at the moment, but we are collaborating with two groups of Unito on these topics.

Title: Accessibility of Internet of Things Devices and Systems
Proposer: Fulvio Corno, Luigi De Russis
Group website: http://elite.polito.it/
Summary of the proposal: The Internet of Things (IoT) is expected to have a significant impact on our daily lives as it will change how we interact with each other, with our objects and our spaces, be they the home or the work place. The IoT provides an opportunity to ensure equal access for everybody: e.g., operating doors and lights through voice makes an IoT-powered environment more accessible to people with physical disabilities and inclusive to many more. Furthermore, web standards for the IoT like the Web of Things extend the open web platform to the physical world and provide a good starting point for making the IoT an enabler, and not a disabler. However, several challenges are still present, ranging from the interoperability of assistive technology with "new" IoT systems to accessibility support in the way we interact with and configure IoT devices, to accessibility guidelines.
This PhD proposal aims at investigating these and similar issues and to provide novel and promising strategies, frameworks, tools, and interface that can positively impact the accessibility of contemporary IoT solutions.
Rsearch objectives and methods: The general objective of this PhD proposal is to assess and improve the level of accessibility of new technologies (based on IoT systems). The application areas of such systems range from smart environments (home, workplace), smart transportation (especially public transportation), health and well-being, gaming and entertainment, as well as socialization and communication solutions. Hundreds of new 'connected' products are constantly being proposed by researchers, startups and large industries. While the security and reliability of many of these solutions is a critical issue, this proposal will mainly focus on their interface and usability, in particular for persons with disabilities (such as speech impairment, cognitive impairments such as autism, motor disabilities).
New smart devices, coupled with novel and intelligent interaction methods, could profoundly empower persons with disabilities in their daily needs. However, many such products are not designed with this need in mind.
One first, minor, research objective is to identify the major accessibility barriers in current IoT systems and devices, both at the level of the adopted interaction methods (voice, touch, etc), and of the available user interfaces. In this, we will identify major issues and shortcomings, and propose alternative solutions (at the algorithm level or at the interface level) for improving their accessibility. The research methods will be based on user studies, controlled experiments, and interaction prototypes.
A second, major, research objective is to propose robust and applicable solutions to some of the challenges identified in the previous phase. In this case, we will research, design, prototype and evaluate a novel IoT device or interface able to solve the accessibility barriers for a specific kind of disabilities. Some examples (that will be validated during the study) are: voice assistants for persons with dysarthria or other speech disorders, alternative interfaces to physical IoT devices for persons with motor disabilities, dynamic generation of user interfaces for IoT devices for persons with mild cognitive disabilities (including the elderly), tangible and immersive IoT-based gaming for children with disabilities, etc.
The research methods will combine the techniques adopted in the Human Computer Interaction field (such as Human centered design processes, empirical evaluations, user studies, etc) with know-how about the design and realization of IoT systems (such as distributed systems, embedded devices, mobile interfaces, intelligent algorithms, etc).
Throughout the doctorate, we expect to activate one or two research projects relevant to these topics.
Outline of work plan: The work plan will be organized according to the following four phases. Phase 1 (months 0-6): literature review about accessibility of digital devices and interfaces; study and knowledge of IoT devices and related communication and programming practices and standards.
Phase 2 (months 6-18): assessment and analysis of accessibility of IoT devices and interfaces. In this phase, some usability assessment and testing, and some accessibility evaluation will be conducted on some significant IoT devices (e.g., conversational assistants, smart thermostats, wearable devices, …) to assess their degree of accessibility for different populations of users. The expected result is the definition of a set of accessibility issues and challenges.
Phase 3 (months 12-24): research, definition and experimentation of novel accessible interfaces for IoT systems. Starting from the results of the previous phase, some of the identified challenges will be addressed in this phase, by proposing new solutions for lowering the accessibility barriers. Such solutions could imply the design of new user interfaces (within the same device or on a different one), or improvement of the intelligent algorithms of the IoT device to compensate for user capabilities.
Phase 4 (months 24-36): the last phase will be twofold. On one hand, the work on addressing the identified challenges will be continued. On the other hand, a robust and large-scale accessible IoT system will be proposed, that includes the proposed technologies and interaction methods, and will be the basis for a comparative assessment of the attained accessibility improvement.
Expected target publications: IEEE Internet of Things Journal, ACM Transactions on Internet of Things, ACM Transactions on Accessible Computing, IEEE Pervasive Computing, IEEE Transactions on Human-Machine Systems, ACM Transactions on Computer-Human Interaction
Required skills and competences:
Current funded projects of the proposer related to the proposal: One VivoMeglio grant (Fondazione CRT). Submitted a project on autism to the TIM Foundation.
Possibly involved industries/companies:None at the moment

Title: Security of "smart" environments
Proposer: Antonio Lioy + Fulvio Corno
Group website: http://security.polito.it/
Summary of the proposal: Society is increasingly going in the direction of “smart” environments (e.g. smart-home, smart-city) that are based on a mixture of IT technologies (IoT devices, cloud computing back-ends, machine learning and big data algorithms). However these technologies and their interconnection present serious security issues that could have severe consequences for people, including physical harm.
As recent news have highlighted, many IoT systems use insecure protocols, expose weakly protected services, don't allow users to secure them, or run on obsolete operating systems. The issue of IoT security is a major concern, as more and more sensitive data are generated and/or handled by these devices and they are becoming part of important architectures (e.g. smart cities).
IoT devices used in healthcare could pose threats to human lives, as would do the devices used in the monitoring and control of critical infrastructures (e.g. water or power supply, train or airplane traffic control).
The PhD candidate will explore all these issues, assess vulnerabilities, and propose possible hardening solutions and best practices.
Rsearch objectives and methods: The PhD objective will be to identify security threats and vulnerability on the design of IoT systems and of IoT devices. In particular, specific solutions will be proposed for the hardening of such systems, for the detection of security risks, and for a robust configuration by the end users.
Research methods will include:
- the study of protocols adopted in wireless networks (from the point of view of their security), and of the APIs provided by middleware gateways and by cloud services dedicated to IoT integration.
- the experimental analysis and security assessment of real-life devices and system
- the analysis of cryptographic techniques viable for IoT devices (that are typically limited in CPU speed and power, memory, communication speed, and electric power)
- the proposal of new communication paradigms and the development of management solutions aimed at containing the security issues based on formal specification and deployment of security policies
- the exploration and proposal of techniques to protect and monitor cloud infrastructures, based on recent advanced network security standards (e.g. MACsec, DNS-SEC, DANE) and integrity verification techniques (the Trusted Computing paradigm and the Trust Zone or SGX elements available on recent CPUs)
Outline of work plan: Phase 1 (months 1-12): study of IoT systems, cloud computing, and ICT security, with special reference to wireless communications, devices with limited capabilities, and Trusted Computing techniques.
Phase 2 (months 6-18): security assessment of smart environments, from the points of view of the devices, of the adopted protocols (both for the local communications and towards the cloud), of middleware gateways (typically used as a local management point and as a front-end towards the Internet), and of cloud services (used for integration of devices, data aggregation and analysis).
Phase 3 (months 18-30): research proposals of hardening methodologies and their individual experimental validation.
Phase 4 (months 24-36): integration of the proposals into a coherent security architecture for smart environments and experimental evaluation of the integrated solution.
Expected target publications: IEEE Internet of Things Journal
IEEE Transactions on Human-Machine Systems
ACM Transactions on Cyber-Physical Systems
IEEE Transactions on Secure and Dependable Computing
IEEE Security and Privacy
International Journal of Information Security
Computers and Security
Software Practice and Experience
Required skills and competences:
Current funded projects of the proposer related to the proposal: SHIELD (H2020) Cybersecurity4Europe (H2020)
Possibly involved industries/companies:Telecom Italia Reply Consoft Sistemi LINKS foundation

Title: Digital Wellbeing in the Internet of Things
Proposer: Luigi De Russis
Group website: https://elite.polito.it
Summary of the proposal: Developers, tech industries, and researchers are designing and creating apps for achieving “digital wellbeing” with smartphone. However, as other smart devices (e.g., smart TVs, smart speakers, …) move into our homes and workplaces, we risk being overwhelmed by a greater amount of distractions, able to further undermine our wellbeing and our control. In addition, people typically use more than one type of device, and use them in a concurrent way, moving from their smartphones to tablets to laptops (and back), thus further complicating this scenario. More work should be put into understanding, tackling, and evaluating multi-device Internet of Things (IoT) ecosystems to support and enhance digital wellbeing. While important initiatives like https://wellbeing.google start to consider multiple services, they mainly use a smartphone-centric approach and do not differentiate among different devices and contexts of use, e.g., people or living places. Furthermore, the approaches currently adopted for smartphones are not particularly effective. This PhD proposals aims at investigating these and similar issues and at providing novel and promising strategies, frameworks, tools, and interfaces that can successfully help self-manage our digital technologies within diverse IoT ecosystems.
Rsearch objectives and methods: According to a 2016 survey, 57% of people use more than one type of device, while 21% are concurrent users of multiple devices. In the last three years, such a joint and multiple usage could have been increased, due to the spread of hundreds of “new” connected products, with novel and intelligent interaction methods, which are constantly proposed by researchers, startups, and large industries. As such smart devices move into our environments, we risk being overwhelmed by a great amount of distractions that harm our digital wellbeing. While recent work focuses on digital wellbeing in the smartphone context, no or few works consider more complex settings.

This PhD proposal aims at closing this gap by answering two main research questions: 1) How can we understand and define which is the impact on our wellbeing of connected and smart devices in a multi-device ecosystem? 2) How can we design framework and devices to develop applications or systems that address the negative impact and increase any positive outcome?

Since the literature on digital wellbeing in the smartphone context questioned the effectiveness of current approaches based on blockers and timers, a simple extension of those approaches in such a more complex ecosystem is not desirable. The main objective of this proposal is, therefore, to define, understand, and design both digital wellbeing strategies suitable for different existent IoT contexts and to propose new technical solutions (e.g., set of ad-hoc IoT devices and/or new interaction modalities) that are “calm” enough and able to fully exploit the possibilities of a multi-device setting.
Digital wellbeing strategies and related technical solutions will be implemented in dedicated frameworks, systems, and tools, and evaluated with target users, possibly in multi-device IoT ecosystems.

A particular attention will be devoted to those ecosystems, as well. While their application areas may range from smart environments (home, workplace, city) to health and entertainment, an important part of the research objective is to explore which kinds of smart and interactive device we can design for a concurrent, yet calm, usage.

For instance, the PhD student may work on new paradigms and technical solutions to detect and change negative multi-device digital “habits” (i.e., recurrent patterns of devices’ usage) at home, through dedicated browser extensions, mobile apps, add-ons for smart TVs and speakers, ad-hoc smart devices, etc., able to share information among themselves and with their users, through dedicated framework with orchestration capabilities. The detection of multi-device habits from usage data will employ a data analytic methodology (e.g., based on association rules), while a promising strategy to selectively change some habits is implementation intentions, i.e., “if-then” plans where ifs are contextual cues and thens are specific goal-related and user-defined behaviors for changing that habit. An example a user may choose to define this: “if it is 6-9pm and I am at home, go for a brief walk (instead of using Facebook on the tablet)” as one of her implementation intentions.
Some of those implementation intentions can be appropriate for some devices, only, e.g., for the user’s smartphone but not for their smart speaker, while others can be shared among different devices. Therefore, a suitable framework or system for exchanging data as well as jointly detecting and merging multi-device habits must also be taken into account.

Finally, the research methods will combine techniques stemming from Human Computer Interaction with know-how about the design and realization of IoT systems (such as distributed systems, embedded devices, mobile interfaces, intelligent algorithms).
Outline of work plan: The work plan will be organized according to the following four phases, partially overlapped.
Phase 1 (months 0-6): literature review about digital wellbeing in various contexts; study and knowledge of IoT devices and smart appliances, as well as related communication and programming practices and standards.
Phase 2 (months 6-18): based on the results of the previous phase, definitions and development of a set of use cases, interesting contexts, and promising strategies to be adopted. Initial data collection for validating the identified strategies in some contexts and use cases.
Phase 3 (months 12-24): research, definition, and experimentation of multi-device technical solutions for digital wellbeing, starting from the outcome of the previous phase. Such solutions will likely imply the design, implementation, and controlled evaluation of distributed and intelligent systems, able to take into account both users’ preferences and capabilities of a set of connected devices.
Phase 4 (months 24-36): extension and possible generalization of the previous phase to include different contexts and use cases. Evaluation in real settings over long period of times to assess at which extent the proposed solutions are able to address the negative impact and increase any positive outcome on our digital wellbeing.
Expected target publications: IEEE Internet of Things Journal, ACM Transactions on Internet of Things, IEEE Transactions on Human-Machine Systems, ACM Transactions on Computer-Human Interaction, ACM Transactions on Cyber-Physical Systems, International Journal of Human Computer Studies, ACM CHI, ACM Ubicomp, ACM CSCW
Required skills and competences:
Current funded projects of the proposer related to the proposal: None at the moment
Possibly involved industries/companies:None at the moment

Title: Evaluating the impact of automated decision systems in urban contexts
Proposer: Juan Carlos De Martin, Francesca Governa, Antonio Vetrò
Group website: https://nexa.polito.it/
Summary of the proposal: Nowadays automated decision (and recommendation) systems are used in several domains such as penal justice, advertising, credit scoring. They take or suggest significantly impacting decisions on the life of people, and it has been widely reported that these tools can cause discriminations against certain population groups (e.g., people with dark skin, females, etc), due to bias incorporated in the datasets used by these tools, or in the algorithms themselves [1]. While law is still failing in regulating the usage of automated decision systems, journalists and researchers collected relevant evidence about the problem, however consensus among mitigation strategies is still far from being reached.
Among the cases reported in the literature, many concern applications of such systems in urban contexts, where physical spaces and characteristics of population influence each other [2]. The PhD candidate will investigate automated decision systems in urban context to understand whether they amplify or reduce existing inequalities in cities, or they create new ones.

[1] O’Neil, C. (2017). Weapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy . New York: Broadway Books.

[2] Wacquant L. (2016), Revisiting territories of relegation: Class, ethnicity and state in the making of advanced marginality, Urban Studies, 53(6) 1077–1088.
Rsearch objectives and methods: The PhD proposal is in collaboration with the Future Urban Legacy Lab. The general goal of the PhD proposal is to understand whether automated decision systems in urban context amplify or reduce existing inequalities in cities, or they create new ones.

The objectives are the following ones:
O1: Understanding inequalities in cities with an interdisciplinary approach A preliminary step in the research is to study inequalities in the urban field, understanding influencing factors, how it is measured, how it is represented. These are aspects of interdisciplinary research that will require the skills and knowledge of those who deal with urban inequality in geography and in socio-economic studies at the Future Urban Legacy Lab. Research on effective graphical interfaces could also require interaction with experts in design and human computer interaction. In addition, the study requires a strong interaction with the scholars of the Nexa Center, who deal with the same subject from the philosophical, political, sociological point of view, and with techniques from the discipline of empirical software engineering.
O2: Understanding the relationship between inequalities in cities and the usage of automated decision systems
This objective should be reached with a two-fold approach:
i) an analysis of scientific evidence and journalistic investigations to understand the current documented impact of automated decision systems in urban context (e.g., digital welfare, predictive policing, risk scores)
ii) experimentation/simulation with real/synthetic dataset
O3: Design, implement and test alternative tools
The PhD candidate should design, implement and test remediation strategies to the main issues collected in previous step. The main expected outcome is a quantitative analysis that shows how the proposed alternative algorithms or data curation processes techniques lessen the disparate impact of classification and/or prediction tasks. In addition, the candidate shall elaborate qualitative insights and policy suggestions on how to shape the digital infrastructure and the processes for decision making software, and to elaborate scenarios on different domains. It will be given the possibility to interact with different institutions and public administrations that are already current interlocutors of FULL and of the Nexa Center (e.g. City of Turin, Lombardy Region, Agency for Digital Italy).
Outline of work plan: - O1: Understanding inequalities in cities with an interdisciplinary approach
o A1.1 Elaboration of a conceptual and operational data measurement framework to measure inequality in the urban field
o A1.2: Definition of a set of guidelines for visualizing inequalities in the urban field
o A1.3: Implementation of a few examples of urban inequality maps following guidelines
- O2: Understanding the relationship between inequalities in cities and the usage of automated decision systems
o A2.1 Literature review on the impact of automated decision systems in urban context
o A2.2 Elaboration of a conceptual and operational data measurement framework for identifying data input characteristics that potentially affect the risks of discriminating software decisions.
o A2.3 Experimentation with (or simulation) of an existing (or hypothetical) automated decision system, which implies classification or prediction tasks on specific population groups, and analysis of impact, also in relation to different fairness metrics already established in the scientific literature
- A3: Mitigation strategies
o A3.1 Based on previous activities findings, design and implementation of mitigation and remediation strategies to reduce the negative impact of classification and/or prediction tasks.
o A3.2 Complement with explanations and critical reflections in the context of digital infrastructure for the future of cities, in relation to ongoing activities at the Future Urban Legacy Lab.
Expected target publications: Illustrative examples of targeted scholarly journals include:

Transactions on Software Engineering and Methodology
ACM Transactions on Information Systems
Expert Systems with Applications
Government Information Quarterly
Journal of Systems and Software
Software Quality Journal
Big Data and Society
Big Data Research
Significance
Daedalus
Information sciences
IEEE Access
IEEE Computer
IEEE Software
Communications ACM
Required skills and competences:
Current funded projects of the proposer related to the proposal: -
Possibly involved industries/companies:Potential involvement of the following industries/organizations:

TIM/Olivetti
La Stampa
Agenzia per l’Italia Digitale
ISO/IEC (Joint Technical Commission 1/ SC 7 (Software Engineering) / WG6 ( Software product and system quality)

Title: Deep Natural Language Processing in Cross-lingual Domains
Proposer: Luca Cagliero
Group website: https://dbdmg.polito.it
Summary of the proposal: Deep Natural Language Processing (NLP) focuses on applying both representation learning and deep learning to tackle NLP tasks. The main goal is the analysis of text morphology, syntax, and semantics to extract relevant knowledge. Examples of applications are machine translation (i.e., translate a text from one language to another), sentiment analysis (i.e., extract subjective information from a text such as opinions, moods, rates, and feelings), and question answering.

In recent years, many vector-based representations of text have been proposed (e.g., Word2Vect, GloVe, FastText, BERT). They analyze the context of use of a text snippet to capture its semantic similarity with other snippets. For example, words in a vocabulary can be represented in a latent space based on the pre-trained models generated from large sets of documents (e.g., the English-written Wikipedia collection). Although the use of Deep NLP models for analyzing English-written documents is established, their portability to contexts in which cross-lingual data need to be analyzed is still limited.

The candidate will investigate and extend the state-of-the-art Deep NLP models. The aim is to propose new solutions tailored to the most challenging cross-lingual domains, e.g., cross-lingual sentiment analysis, query-based text summarization of multilingual documents, opinion mining from cross-lingual social data.
Rsearch objectives and methods: The research objectives are enumerated below.
- Application and integration of existing Deep NLP models. Study the applicability of state-of-the-art Deep NLP models to challenging research contexts, such as document summarization, sentiment analysis, and anomaly detection from text.
- Retrofitting of Deep NLP models. Tailor the Deep NLP models trained on general-purpose textual documents (e.g., the Wikipedia corpus) to specific domains (e.g., finance, medicine) by using lexical relational resources to obtain higher quality semantic vectors.
- Cross-lingual model alignment and knowledge propagation. Extend Deep NLP models tailored to a source language by integrating knowledge from other languages. Develop new strategies to align multilingual vector spaces and to propagate embedded information across multiple languages and domains.
- Application of cross-lingual models to existing application contexts. Develop cross-lingual solutions to challenging text mining problems based on Deep NLP models. The following application contexts will be considered: document summarization, sentiment analysis, and anomaly detection.

Address the following key issues.
- Integration of multi-lingual embeddings at the word- and sentence-level. The vector-based representations of the documents written in different languages need to be compared with each other and aligned in order to generate a unified model suitable for cross-lingual text analyses.
- Study and development of new cross-lingual models. Deep NLP approaches address specific NLP tasks using representation learning and deep learning. Most of the current solutions have been developed and tested on a single language. Integrating models trained in different languages and using lexical relational resources to obtain higher quality semantic vectors (i.e., retrofitting) are still open problems in the cross-lingual domain. Hence, the aim is to advance the state-of-the-art approaches and extend their portability towards other languages and contexts.
- Investigation of new NLP challenges The recent advances of Deep Learning techniques have enabled researchers to address new text mining challenges. The goal is to investigate and address cutting-edge research challenges with the eye to their portability towards cross-lingual domains. For example, topic detection methods are still very limited to specific contexts and, often, do not allow coping with cross-lingual documents.

The research will address the study, development, and testing of various algorithms and frameworks. Extensions of existing solutions tailored to specific types of documents (e.g. news articles, tweets, scientific articles) will be initially studied. Then, their portability to other contexts (e.g., social data) will be investigated as well.
Outline of work plan: PHASE I (1st year): overview of the existing Deep Learning architectures, study of the state-of-the-art embedding models, analysis of their portability to different languages, and evaluation of different vector alignment strategies. Study of the state-of-the-art NLP solutions to tackle the most common tasks, among which question answering, analogy detection, prediction of the current word from a window of surrounding context words (i.e., the CBOW architecture), and prediction of the surrounding window of context words (i.e., the Skip-gram architecture).
PHASE II (2nd year): study and development of Deep NLP solutions to address existing, open problems. Application of the proposed models to multi-lingual data acquired in various domains (e.g., insurance, learning, finance). Study of existing topic discovery techniques and analysis of their portability to documents written in different languages.
PHASE III (3rd year): Investigation of new, challenging NLP tasks. Extension of the preliminary solutions towards cross/lingual domains. Study and development of new topic discovery techniques oriented to multi-lingual collections. Study of the portability of the solutions designed in the current and previous phases to heterogenous data (e.g., social data).
During all the three years the candidate will have the opportunity to attend top quality conferences, to collaborate with researchers from different countries, and to participate to competitions on data summarization organized by renowned entities and previously addressed by the proponent, i.e., the MultiLing Community Challenges and the Computational Linguistics Scientific Document Summarization Shared Task (CL-SciSumm) overseen by the National University of Singapore.
Expected target publications: Any of the following journals on Data Mining and Knowledge Discovery (KDD) or Data Analytics:

IEEE TKDE (Trans. on Knowledge and Data Engineering)
IEEE TPAMI (Trans. on Pattern Analysis and Machine Intelligence)
ACM TKDD (Trans. on Knowledge Discovery from Data)
ACM TIST (Trans. on Intelligent Systems and Technology)


IEEE/ACM International Conferences on Data mining and Data Analytics (e.g., IEEE ICDM, ACM SIGMOD, IEEE ICDE, ACM KDD, ACM SIGIR)
Required skills and competences:
Current funded projects of the proposer related to the proposal: - Consulting and research contract with Reale Mutua Assicurazioni (Status: ongoing). The activities will be focused on the analysis of textual data coming open data sources and social platforms related to, amongst other, the monitoring of traffic and accident warnings in real time, the conditions of the road routes, and the meteorological alerts.
- PRIN Italian Project Bando 2017 "Entrepreneurs As Scientists: When and How Start-ups Benefit from A Scientific Approach to Decision Making" (Status: Accepted, not yet started). The activity will be focused on the analysis of questionnaires, reviews, and technical reports related to training activities.
Possibly involved industries/companies:

Title: From customer's devices to the cloud: orchestrating the computing continuum
Proposer: Fulvio Risso
Group website: http://netgroup.polito.It
Summary of the proposal: Recent technologies such as edge and fog computing have complemented cloud computing offers with the possibility to share resources at the edge of the network. In the near feature we expect to include also customer devices (laptop, smartphones, and even embedded and IoT devices) in this computing continuum, offering endless locations and unprecedented agility to computing tasks.
On the other side, this paradigm poses non-trivial challenges. In addition to the necessity of scheduling each task in the optimal location (which can change rapidly over time, particularly in case of mobile devices), the above resources are expected to be under the control of different administrative organizations (e.g., production factory, telco edge POP, cloud data center), each one with different goals and optimization functions. In addition, also customers (e.g., Over-The-Top operators), which represent mainly resource consumers, have their objectives which may collide with the resource providers. However, it is well understood that the widespread acceptance of fog/edge computing (and even device-based resource sharing) will happen only if a "win-win" solution will be defined that satisfies all the involved actors.
Given the complexity of this scenario, the current PhD is oriented to investigate the above problem focusing on the case of multi-administrative actors, in particular (1) when telcos are active part in this game and (2) when economic considerations are included in this picture.
Rsearch objectives and methods: The candidate could pursue one (or more) of the following directions to reach the objective, while guaranteeing the "win-win" property mentioned above and the necessity to establish economic relationships among parties:
- Paradigms, scalable algorithms, and protocols to advertise, negotiate and acquire resources in another administrative domain, with the guarantee that advertised resources and negotiated price would be shared only between the involved parties;
- Scalable algorithms and protocols to monitor the computing/networking infrastructure, in order to evaluate, in real-time, the actual convenience of offloading a task in a foreign domain;
- Scalable state-sharing mechanisms that enable a hosting domain to keep control of its own resources (even if offered to a foreign actor), while allowing the customer to monitor the state of the remote jobs such as they were actually running in its own administrative domain;
- Effective and dynamic network virtualization technologies that enable to extend a given administrative domain to include "foreign" resources, which are all perceived such as local.

Finally, a use case will be defined in order to validate the above finding in more realistic conditions. Among the possible choices, OTT delivering services across different domains (e.g., cloud datacenter, edge telco cloud, local infrastructure such as stadium), enterprise IT services (spanning across the enterprise network but including also telco and other third parties infrastructures), and more.
Outline of work plan: The proposed research plan, which covers a subset of the possible objectives listed in the previous section, is structured as follows (in months):
- [1-5] State of the art
o Scheduling and job allocation
o Mathematical foundation: game theory / auctions
- [6] Paper writing: survey and state-of-the-art of distributed task scheduling
- [7] Extension of a cloud toolkit to include also end-devices (e.g., through Rancher K3s).
- [8-10] Design and implementation of a scheduler for intra-domain jobs, including quality-related parameters such as latency, bandwidth, expected response time.
- [11-14] Validation through simulations (for large-scale testbeds) and by physical setup, with the collection of real metrics.
- [15] Paper writing (conference).
- [16-21] The multidomain case: extension of the scheduling algorithm for a multidomain scenario, in which administrative policies (in additional to physical constraints) play a fundamental role.
- [22-24] Validation through simulations (for large-scale testbeds) and by physical setup, with the collection of real metrics.
- [25-26] Paper writing (journal).
- [25-36] In parallel, start a low-intensity task oriented to create and guide a working group in an existing open-source cloud orchestration toolkit (e.g., Kubernetes) in order to create a standard for multi-domain interconnections (protocols, API), mimicking what the Border Gateway Protocol (BGP) does in case of multi-domain networks. Possibly this should generate a publication focuses on the defined protocol and programming interface.
- [27-31] Large scale setup and monitoring of run-time data with the help of a use-case partner (possible choices: (1) factory, (2) hospital).
- [31-34] Possible evolutions of the algorithms; push of the solution toward open-source software; paper writing (magazine).
- [35-36] Writing PhD dissertation.
Expected target publications: Top conferences:
- USENIX Symposium on Operating Systems Design and Implementation (OSDI)
- USENIX Symposium on Networked Systems Design and Implementation (NSDI)
- International Conference on Computer Communications (INFOCOM)

Journals:
- IEEE/ACM Transactions on Networking
- IEEE Transactions on Computers
- ACM Transactions on Computer Systems (TOCS)

Magazines:
- IEEE Computer
Required skills and competences:
Current funded projects of the proposer related to the proposal: None
Possibly involved industries/companies:Telecom Italia Mobile, TOPIX (Torino Piemonte Internet Exchange)

Title: Planning Safety in the Era of Approximate Computing Systems
Proposer: Stefano Di Carlo
Group website: http://www.testgroup.polito.it
Summary of the proposal: Cyber-Physical Systems (CPSs) are the root of the fourth industrial revolution and the Internet of Things (IoT) forms a foundation for this cyber-physical systems revolution. CPS are large physical systems that consist of many interacting physical elements and distributed IT systems for monitoring, control, optimization and interaction with human operators and managers. All devices of a CPS and different CPSs are connected in order to create a network enabling billions of systems and devices to interact and share information.
While CPSs are crucial to many application domains, several modern computations (such as video and audio encoders, Monte Carlo simulations, and machine learning algorithms) are designed to trade off accuracy in return for increased performance and lower power consumption. To date, such computations typically use ad-hoc, domain-specific techniques developed specifically for the computation at hand and none of them consider the final safety impact on the CPSs.
In this context, this Research Proposal is particularly interested at analyzing the safety of CPSs when approximate computing is used. The main goal is to analyze several approximation paradigms when dealing with safety constraints and to propose alternative strategies to cope with any safety issue.
Rsearch objectives and methods: The research objectives envisioned for this work include both approximate computing and safety aspects:

1. To develop approximate computing methods to assess the impact on safety. Those methods will properly guide the application of approximation under safety constraints. The Safety standards should come from the most interesting fields of application, such as automotive and aerospace.
2. To study and to propose Safety-oriented approximate computing models to target an early analysis of the system under approximation usage. Those models must enhance the capability of early designing approximated CPS under safety constrains by playing with all system parameters. The parameters should include all layers of a CPS, such as hardware components and their specific instantiation (e.g., CPUs, GPUs, FPGA, etc.) as well as middleware (e.g., Real-Time Operating systems) and software.
3. To provide tools to carefully design CPS including the approximate computing without failing safety constraints. Such tools should include design space exploration techniques to balance the larger number of design constraints, such as power consumption, area, time, etc. Previously developed methods and models should empower those tools with the necessary flexibility in meeting the constrains and system final design.

Those objectives could also lead to the participation or proposition of new Safety standards.
Outline of work plan: Year 1: Safety standards and approximate computing estimation methods

The first step to achieve the goals described in the previous section is to analyze all current approximate computing strategies alongside with the current Safety standards to develop effective estimation methods to properly link the two worlds.

To carry out this activity the student requires to acquire a deep knowledge of the state-of-the-art in order to identify safety-related case-studies and AxC techniques to be applied. The estimation methods should evaluate the accuracy and the safety as well as the power, performance.

Year 2: From Methods to Models and Paradigms

The second step carries out the transition from the estimation methods to the proper modeling of the CPS system in order to ease the analysis of the safety impact. The student must become familiar with several modeling, programming and simulation frameworks at different layers. Artificial Intelligence models and Machine Learning techniques could be used to properly analyze the models, and the definition of new paradigms for the evaluation can be part of this step of the work.

Year 3: Tools, Design Space Exploration and further model improvements

The last year of the Ph.D. will be dedicated to the development of tools and DSE techniques for optimizing CFPs under safety constrains. These approaches will exploit the capability of the models developed during the second year of the Ph.D. project coupled with efficient space exploration algorithms for multi-objective optimization such as the extremal optimization theory or other evolutive approaches.

The DSE will be fed with multi-objective goals in order to provide an instrument to trade-off safety and accuracy with power and performance.
Expected target publications: The work developed within this project can be submitted to:

Conferences (DSN, DATE, DAC, IOLTS, etc.): a minimum of two conference publications per year.

Journals: IEEE TOC, IEEE TCAD, ACM TACO, IEEE TDSC: at least one publication during the second year and one during the third year.
Required skills and competences:
Current funded projects of the proposer related to the proposal: Intel Corporate Italia, Leonardo.
Possibly involved industries/companies:

Title: Algorithms and technologies for ubiquitous applications
Proposer: Filippo Gandino, Renato Ferrero
Group website: http://www.cad.polito.it/
Summary of the proposal: Ubiquitous computing is an innovative paradigm of human-computer interaction, which aims at the integration of technology into everyday objects to share and process information. Two main characteristics of this vision are accessibility of computing and communication services every time and everywhere, and calm technology, with minimal attention required to the user for the system interaction. Current technology trends are leading from one side towards smart devices, which offers mobile, complex and personalized functionalities, and from the other side towards embedded systems disappearing in the physical world and performing simple and specific tasks. It follows that several issues must be addressed for the development of ubiquitous applications: architecture design for interlinking the ubiquitous components (smart devices and embedded systems); use of a large variety of technology (e.g., sensor networks for monitoring physical conditions, RFID network for tagging and annotating information, actuators for controlling the environment); development of algorithms for smart and autonomous functionalities. The research activity of the PhD candidate will cover all the phases related to the design, development and evaluation of ubiquitous applications, so he/she is required to own multidisciplinary skills (e.g., distributed computing, computer network, advanced programming).
Rsearch objectives and methods: The research objectives concern the identification of requirements and the investigation of solutions for designing and developing ubiquitous applications. More in details, the following topics will be addressed:
1) study of distributed architectures able to support both local and remote services, thus increasing the number of functions offered and avoiding duplication of code. Resource availability and access (storage, applications, data for end-user) will be based on cloud computing and fog computing.
2) enhancement of the desktop computing paradigm, which is based on an explicit human-computer interaction. In fact, as the number of embedded systems dramatically increases in the ubiquitous vision, an explicit interaction with them would distract and overwhelm users. Instead, the ubiquitous system should understand and react to the actions of the user, even if they are not primarily intended as an input to the system. More natural and implicit kinds of interaction, such as natural language, will be evaluated by exploiting for example the available API for natural language processing (such as the ones provided for Google Assistant and Amazon Alexa).
3) support for context awareness. The involved technology consists of wireless sensor networks, which can provide useful information about the physical environment (e.g., location, time, temperature, light…), and RFID networks, for context-based query, item location and tracking, automated access to physical place or virtual service. A context-aware application identifies the most useful services for the user, instead of proposing the full list of available functionalities. This can limit the effort of the user in interfacing the system and can reduce the computational requirements.
4) development of strategies for enhancing the autonomy of the ubiquitous system. For example, algorithms for power savings will increase the lifetime of mobile components, thus reducing the need of human maintenance.
5) development of "smart" systems for proactive behavior and adaptability in dynamic contexts. The research will focus on modeling the physical environment, as well as human behavior. Limitations are due to the dynamicity of the environment, its incomplete observability, the impossibility to completely determine the user actions and goals. Therefore, algorithms for handling the incompleteness of the system and the non-deterministic user behavior will be designed.
Outline of work plan: The PhD research activities are organized in three consecutives phases, each one roughly corresponding to one year in the PhD career.
In this first phase the PhD candidate will improve his/her background by attending PhD courses and by surveying relevant literature, then he/she will apply the learnt concepts to tackle specific issues in the implementation of ubiquitous systems. In particular, the first research hints will regard applications already developed by the research group. For example, thermal monitoring in smart building is currently an active field for the research group and the PhD candidate may be in charge of enhancing the ubiquity of applications developed in this context. As already existing applications will be considered in the first phase, the five research objectives can be considered as independent from each other. In this way it is possible to anticipate the expected outcome (personal scientific contribution, research papers) by focusing on one research objective at a time, since there is no need to master all concepts in advance.
In the second phase, when the training is completed and the PhD candidate owns a full vision of the matter, he/she will be able to design new solutions and propose technologic improvements to problems in the research group expertise, in order to fully support the automation of human tasks and the access to information anywhere and at any time.
In the third phase, the maturity of the PhD candidate will allow to identify and deal with general challenges of ubiquitous computing, such as the obtrusiveness and the human overloading during the system interaction. The PhD candidate will be involved in research projects aiming at designing and implementing new ubiquitous applications. He/she will be required to exploit his/her competence to analyze and solve real problems and finally to evaluate the performance of proposed solutions.
Expected target publications: - IEEE Pervasive Computing
- IEEE Journal on Internet of Things
- ACM International joint Conference on Pervasive and Ubiquitous Computing (UbiComp)
- IEEE International Conference on Pervasive Computing and Communications (PerCom)
Required skills and competences:
Current funded projects of the proposer related to the proposal: TECH-START - key enabling TECHnologies and Smart environmenT in the Age of gReen economy. PRIN 2017 (Filippo Gandino)
Possibly involved industries/companies:None

Title: Pervasive Information Management
Proposer: Filippo Gandino
Group website: http://www.dauin.polito.it/research/research_groups/cad_electroni...
Summary of the proposal: Pervasive technologies are composed by densely distributed, networked, low cost devices. The two main examples are wireless sensor networks and radio frequency identification. This kind of technologies involve many devices that continuously collect, elaborate and transmit data. The proper processing of this information is fundamental to exploit the actual benefits of pervasive technologies. However, this task is very complex and represents one of the main issues that are slowing their diffusion.

The research activity will be related to the whole cycle of life of the pervasive information, starting from the collection, through the transmission, up to the final analysis. The candidate will work within a team with a consolidated experience in pervasive technologies and he/she will have the opportunity to access to implemented and under development sensor networks.
v The proposed research involves multidisciplinary knowledge and skills (e.g., data mining, computer network, advanced programming).
Rsearch objectives and methods: The goals of the research are related to the whole cycle of life of the pervasive information.

The first phase is the data collection. It is characterized by which data are produced, their format and how often they are collected. These characteristics are fundamental, since they affect the accuracy of the system, the power consumption and the quantity of data to store and transmit.

The second phase is the pervasive elaboration. The devices directly elaborate the collected data. They can merge more data, compute statistics, apply filters, correct values. However, the quantity and quality of these tasks must be defined by considering that the computational capability of pervasive devices are often limited like their power supply.

The third phase is the transmission. According to the previous phases, the amount of data is already determined. The frequency of the transmissions will affect the efficiency of the system update, and the power consumption.

The fourth phase is data storage. In order to effectively manage big data, the data base organization is fundamental. Moreover, a part of the data can be deleted or summarized in order to save space and improve the query efficiency.

The fifth and last phase is the data analysis and visualization. The data must be analyzed in order to find useful information. Moreover, the data analysis is important for the accuracy of the system, since the data must be calibrated, even dynamically, odd data must be removed and missing data can be inferred.

The candidate will have the opportunity to work with already working sensor networks and to cooperate in the development of new ones. Its activity will focus on the balance between efficiency and data accuracy and on the data analysis. The goals include solving specific problems and finding general strategies and techniques both in the organization of the data and in their analysis.
Outline of work plan: The candidate will start its activity by becoming familiar with the pervasive systems already developed within the research group. He will analyze the phases of the cycle of live of the information in order to improve the balance between efficiency and accuracy. During the phd, he/she will also have the opportunity to work on the design and development of new pervasive systems.
In parallel, a large part of the research activity will be focused on the last phase: the data analysis.
A relevant topic is the calibration of the data. Although there are consolidated techniques for the static calibration, they are not always applicable. The candidate will study how to calibrate data when the standard solutions are not enough. Moreover, he/she will investigate the technique for dynamic calibration, in which when two sensors meet, their calibration is updated. Starting from the state-of-the-art work, the activity will consider how to improve the accuracy also of older values according to the new calibration.
A second topic is the filtering of the data in order to delete unreliable data or to tag them as partially reliable. This task will be executed by investigating the correlation between the reliability of the data and other values (e.g., speed, humidity).
A third topic will be the inference of the missing data. This task is fundamental for data with a bi- or three-dimensional position. Moreover, with mobile sensors, not only a spatial inference but also time inference could improve the results. The candidate will investigate this topic, and will use the results also for the optimization of the node deployment.
The forth topic is the analysis of the data. Each sensor system provides different opportunities that can be reached only through effective analysis of the data. The candidate will exploit data mining and machine learning approaches.
Expected target publications: IEEE Transactions on Parallel and Distributed Systems
IEEE Transactions on Wireless Communications
IEEE Internet of Things Journal
Wirekless Networks (Springer)
Required skills and competences:
Current funded projects of the proposer related to the proposal: TECH-START - key enabling TECHnologies and Smart environmenT in the Age of gReen economy. PRIN 2017
Possibly involved industries/companies:No

Title: Key management techniques in Wireless Sensor Networks
Proposer: Filippo Gandino
Group website: http://www.dauin.polito.it/research/research_groups/cad_electroni...
Summary of the proposal: Wireless sensor networks (WSNs) offer benefits in several applications but are vulnerable to various security threats, such as eavesdropping and hardware tampering. In order to reach secure communications among nodes, many approaches employ symmetric encryption. Several key management schemes have been proposed in order to establish symmetric keys exploiting different techniques (e.g., random distribution of secret material and transitory master key). According to the different applications of WSNs, the state-of-the-art protocols have different characteristics and different gaps.

The proposed research activity will be focused on the study and development of key management protocols. The proposed investigation will consider both the security requirements and the performance (e.g., power consumption, quantity of additional messages, computational effort) of the networks. The research will analyze different possible solutions, evaluating the trade-off in terms of costs and benefits, according to different possible scenarios and applications.

The initial part of the activity will be focused on the proposals currently in progress within the research group. However, the candidate will be stimulated to formulate and develop new solutions.

The proposed research involves multidisciplinary knowledge and skills (e.g., computer network and security, advanced programming).
Rsearch objectives and methods: The objectives of the proposed activity consist in studying the state-of-the-art of key management protocols for WSNs and to propose new solutions suitable to various application scenarios. The requirements that affect the security protocols will be considered in order to find the best solution for different kinds of WSNs. In particular, the mobility of the nodes and the possibility to add new nodes after the first deployment will be considered.

According to the current trends in this research field, the starting activity will be focused on the use of combinatorial approaches for the distribution of the keys. The use of these techniques is considered appealing, since the provided security properties are more effective than the ones based on randomness. However, they involve drawbacks in terms of memory overheads and limits to the size of the network.

Moreover, the research activity will consider key management in network dedicated to specific applications. In particular, the first considered case study will be the Internet of Things in which WSNs represent an important building block and require specific solutions for an effective interoperability.

The candidate will start to cooperate to the ongoing research activity by implementing, testing and evolving the proposals actually under consideration by the research group. However, a fundamental goal of the candidate is to reach skills suitable for the development of autonomous proposals.

The proposed solutions will be evaluated and compared to state-of-the-art approaches, in order to evaluate their security level and their performance. The analysis of the protocols will be composed by: (a) a theoretical analysis, (b) simulations, and (c) implementation on real nodes.

The theoretical analysis will consider several aspects of the key management protocols. A part of this analysis will evaluate the statistical characteristics of the scheme, in order to analyze the level of security and other network parameters for protocols based on stochastic elements. A security analysis will compare the properties and the requirements of the protocols.

The simulations will be used to analyze the performance of the schemes with a high number of nodes. This kind of investigation allows reaching significant results within a limited time. There are many available tools that allow simulating WSNs, e.g., TOSSIM, which uses the same code written for the TinyOS platform.

The last kind of analysis will be based on an experimental session on a real network. The main platform that will be used will be TinyOS, in order to develop a code that can be used also for the simulations. The main purpose of a real implementation is to validate the results achieved by simulations.

In order to work on this research topic, a candidate may have good programming skills, and he/she should have a good knowledge on networking and security.
Outline of work plan: During the first part of the first year the candidate will investigate the field of the research activity. Therefore, the candidate will study the existing protocols and will improve his/her knowledge on the related topics. This phase lasts 3 months. During the rest of the first year, the candidate will start to work on research proposals currently under consideration within the research group. The PhD student will cooperate in the detailed definition of the key management schemes, and in their implementation, considering a real Internet of Things case study WSN. The preliminary results will be submitted to international conferences.

During the second year, an accurate analysis and evaluation of the proposed ideas will be done, and an eventual modification of those approaches will be considered. This task will include a programming activity, in order to test the proposed solutions and to integrate them in real applications. The achieved results will be submitted to an international journal.

During the last year, the main goal of the candidate will be to conclude the analysis on his proposals and to publish the results. However, also the task of evolution of the current proposals of the research group will be curried on, in order to differentiate its research activity and to increase the probability of reaching valuable results.

During all the three years the candidate will have the opportunity to cooperate in the development and implementation of key management solutions applied to research projects. These tasks will allow the candidate to understand the real scenarios in which WSNs are used and to find their actual issues.
Expected target publications: IEEE Transactions on Information Forensics and Security
IEEE Transactions on Parallel and Distributed Systems
IEEE Transactions on Wireless Communications
Wireless Networks (Springer)
Required skills and competences:
Current funded projects of the proposer related to the proposal: TECH-START - key enabling TECHnologies and Smart environmenT in the Age of gReen economy. PRIN 2017
Possibly involved industries/companies:No

Title: ICT for Urban Sustainability
Proposer: Maurizio Rebaudengo
Group website: http://www.cad.polito.it/
Summary of the proposal: Air quality nowadays is receiving ever growing attention since it has become a critical issue, in fact long-term exposure to polluted air can result in permanent health issues. The causes of air pollution are various: fossil fuels, power plants, factories, waste incineration, controlled burn, cattle farming and so on. All these sources contribute to release several toxic agents. In fact, they can put in danger not only the people - it is one of the major cause of deaths per year - but also our entire ecosystem. The research will investigate a smart city system to help local government and citizens themselves to monitor what currently happens in the city. The designed architecture of the system will use sensor networks able to capture city condition like temperature, air pollution, traffic situation, etc. The system will be deployed in a mobile and static network installed on different entities, like the bicycles available in the bike-sharing service, the public transportation vehicles and fixed spots to acquire real-time and diffused information on the urban area. The huge amount of data will be analyzed to develop mobile applications useful to help the citizens and the stakeholders in the management of the actions to reduce the air pollution or the traffic congestion, and optimizing the use of the urban resources. The research activity will be focused on some open issues related to the efficient application of mobile wireless sensor networks: e.g., the evaluation of the mobile systems in case of deterministic and stochastic routes, the auto-calibration of data collected by low cost sensors, the energy saving based on optimization of transmission and data collection protocols, the application of machine learning approaches to selectively discard erroneous or not relevant data.
Rsearch objectives and methods: The research will study and develop a hardware system and a software tool able to monitor, control and handle urban problems such as air pollution, traffic congestion, water quality, and so on.

The research will design the whole architecture of the system composed of different types of nodes, fixed and mobile. Mobile nodes could be deployed into public transport vehicles and public bicycles. Fixed nodes will be installed in some critical points, like the most crowded crossroads. The use of mobile nodes requires a study on the spatial and temporal coverage due to the possible route strategies (e.g., stochastic routes for bike sharing, fixed paths for busses), in order to verify the trade-off between costs and benefits for an efficient urban air quality monitoring.

Among the amount of hazardous gases, some pollutants (e.g., CO, ground level O3, Particulate Matter, and Pb) are the most dangerous and will be considered as the starting point for the development of the sensor nodes. The use of low cost sensors involves a temporal variation of the sensor output. In order to address this problem, algorithms for the auto-calibration of the data will be studied and developed. In case of mobile node, the localization of each node becomes an essential constraint, in order to realize an ubiquitous real-time monitoring system: the integration of a GPS module, which delivers accurate position and time information, will be considered in order to guarantee a pervasive system. A particular attention will be paid in evaluating the scalability of the system, in order to maximize the number of possible nodes. Possible integration with different technologies will be considered, and in particular RFID-based sensors fixed nodes will be exploited in cooperation with mobile WSN nodes. Data will be exchanged among the nodes and toward a gateway exploiting low-power wireless data routing protocols or a telecommunication network. The research activity will consider the optimization of data acquisition, storing, aggregation and transmission, in order to reduce energy consumption. Data will be processed on the low-level module of the network and then will be collected by distributed gateways and a central server according to a hierarchical architecture that must guarantee their reliability and the availability of the system.

The foreseen activity will consider also the analysis of the huge amount of data with the goal to realize a platform of data acquisition and evaluation, useful to develop mobile applications to be installed in portable or automotive devices, with the goal to display a dashboard containing useful information such as air pollution, traffic condition or parking availability, and to elaborate those information to make some decision useful for the user citizen, e.g., the best path, as far as the traffic is considered, the area with a high or low level of pollution, the zone with available parking, etc.
Outline of work plan: The research activity is organized in 3 main work packages (WPs). The first WP includes the design, implementation and experimentation of heterogeneous devices for monitoring useful parameters. The second WP deals with their integration in a large WSN for a widespread coverage of the city. Finally, the third WP consists in the management and analysis of the data collected from the WSN.

The first activity of WP1 aims to identify, for each urban problem that will be faced, a set of parameters for a complete characterization. For example, traffic congestion can be evaluated recording the number of passages in fixed points, air pollution can be measured according to the level of noxious gases in the air. It is expected that different modules will be integrated to obtain more comprehensive information: for example, a GPS module can be added to a sensor to precisely localize the place of measurement. State-of-the-art protocols for data acquisition, storing, aggregation and transmission will be studied and implemented on the nodes in order to find new optimized energy –saving solutions. In the last part of WP1, the developed devices will be tested on the laboratory and on the field to evaluate their effectiveness in facing the considered urban problems. A special attention will be focused on the auto-calibration of the data, aiming at proposing new solutions in order to improve the accuracy of the collected data.

The activities of WP2 are based on an initial state-of-the-art analysis to identify the most common practical issues in the design of a big WSN, with a special focus on their adoption in smart cities. Existing solutions will be reviewed in order to propose an effective integration of the devices in a robust network. The proposed infrastructure will be implemented and tested. Particular attention will be paid to the efficiency and security of the WSN, in terms of resource consumptions and communication reliability and security. A study on the characteristics of possible vehicles for the implementation of a mobile sensors network will investigate the time and spatial frequency of the measurements, the additional cost (e.g., in terms of human labor) for the system and the possible alteration of the data (e.g., due to the emissions of the vehicle).

The first activity carried on during WP3 is the development of a database for storing the data acquired from the WSN and for efficiently retrieving this information: Data Mining algorithms will be studied and developed for a deep analysis of the measured data; the main activity of WP3 regards the development of dashboard useful to develop solutions exploiting the so-formed knowledge for effectively improving the urban life.

All the results achieved in the WPs will be submitted to international conferences and journals.
Expected target publications: IEEE Transactions on Wireless Communications
IEEE Transactions on Industrial Informaticss
IEEE Internet of Things Journal
Wireless Networks (Springer)
Journal of Network and Computer Applications (Elsevier)
Required skills and competences:
Current funded projects of the proposer related to the proposal: TECH-START - key enabling TECHnologies and Smart environmenT in the Age of gReen economy. PRIN 2017 (Filippo Gandino)
Possibly involved industries/companies:None

Title: Self-evolving data-driven models
Proposer: Elena Baralis, Daniele Apiletti
Group website: http://dbdmg.polito.it
Summary of the proposal: While data-driven models based on data mining techniques and learning approaches are widely exploited in many different application contexts, the analysis of their evolution has only received attention in recent times.
However, ageing of physical components and changes in phenomena under exam often cause drifts in the data over time.
As new data are analyzed by models trained in the past, being able to assess the validity of predictive models is crucial for keeping high-quality results. For instance, detecting the degradation of the currently-in-use models is of paramount importance for triggering an update of the models themselves. Such approach is also a key step towards automating machine learning in real-life contexts.
While in supervised models different metrics exist to capture a degradation in prediction performance, they are often applied in contexts where the ground-truth labels are unavailable, hence opening the research question on how to measure their degradation in an efficient and effective way.
The current state of the art is addressing the challenge, but it is lacking a comprehensive view including the design of innovative quality metrics able to assess the bias between the training set and new data.
The PhD student should design and develop data-driven models and proper evaluation metrics able to automatically detect model degradation over time, under different data velocity, variety, and volume conditions.
Rsearch objectives and methods: Prioritized and progressive research objectives are provided in the following (from the most reachable to the hardest).

1) Data characterization
To understand the structures and the models hidden in a data collection, a set of descriptive metrics will be defined by exploiting unconventional statistical indexes and new algorithms to model the underlying data structures.

2) Model characterization
To identify possible degradation metrics, the models to be evaluated and their learning techniques must be characterized according to their key features over volume, variety, and velocity.

3) Degradation metrics
A set of degradation metrics will be studied and designed to evaluate the performance of the characterized models over time.
Such metrics will take into account the unavailability of ground-truth labels at computation time, the computational costs, and the scalability over large volumes of data, and their applicability to different data types (variety). Incremental computation is also a desired feature.

4) Time horizon
Different time horizons must be addressed, depending on the application context: from near real-time, to long periods, such as those featured by cyclic manufacturing processes with slowly degrading components.

5) Volume, variety, and velocity
Different requirements in terms of volume of the data (i.e., scalability), variety of the data (i.e., heterogeneity), and velocity (i.e., batch vs real-time analysis) are to be addressed, considering the trade-off between more precise yet resource-expensive metrics, and scalable yet approximate approaches.
Outline of work plan: During the 1st year, the student will study state-of-the-art solutions and their limitations. A set of descriptive metrics will be defined to model data distributions. Since several aspects influence data distributions, the PhD student will draft, for each data type (variety), an innovative criterion to model data distribution by exploiting unconventional statistical indexes. Model characterization and time horizon will be also addressed, with initial focus on simpler, less demanding, batch solutions.

During the 2nd year, the PhD student will define innovative unsupervised metrics to evaluate the degradation of predictive models over time. A parallel development and preliminary experimental phase on public, synthetic, and possibly real-world industry-provided datasets is required. The work will also address near real-time solutions and scalability over volume and velocity.

During the 3rd year the candidate will extend the research by widening the experimental session to larger volume, faster velocity, and wider variety datasets, properly adapting the solutions proposed in year 2 to the new challenges. Possibly, the candidate might spent a period (3-6 months) into a foreign university or research center.

The candidate will assess the proposed solutions in real-life application domains based on the candidate interest, research contracts, funded projects, and available datasets. All these application contexts will be key to disseminate the proposed solutions not only in the primary venues (conferences and journals), but also towards industries and the society, fostering cross-domain fertilization of research results, and new collaborations with other research stakeholders.

A planned outcome of the work is to investigate the exploitation of self-evolving models (i.e., concept drift detection techniques) to address the training bias in machine learning.
Expected target publications: Any of the following journals
IEEE TKDE (Trans. on Knowledge and Data Engineering)
ACM TKDD (Trans. on Knowledge Discovery in Data)
ACM TOIS (Trans. on Information Systems)
ACM TOIT (Trans. on Internet Technology)
Information sciences (Elsevier)
Expert systems with Applications (Elsevier)
Engineering Applications of Artificial Intelligence (Elsevier)
Journal of Big Data (Springer)

IEEE/ACM International Conferences and Workshops, EDBT, VLDB
Required skills and competences:
Current funded projects of the proposer related to the proposal: Research contract with Lavazza (resp. Elena Baralis).
Research contract with ENEL (through the SmartData center).
The research contracts with industry allow to focus on real-world use cases with a direct social impact.
Possibly involved industries/companies:None

Title: Promoting Diversity in Evolutionary Algorithms
Proposer: Giovanni Squillero
Group website: http://www.cad.polito.it/
Summary of the proposal: Computational intelligence (CI) in general, and evolutionary computation (EC) in particular, is experiencing a peculiar moment. On the one hand, fewer and fewer scientific papers focus on EC as their main topic; on the other hand, traditional EC techniques are routinely exploited in practical activities that are filed under different labels. Divergence of character, or, more precisely, the lack of it, is widely recognized as the most impairing single problem in the field of EC. While divergence of character is a cornerstone of natural evolution, in EC all candidate solutions eventually crowd the very same areas in the search space, such a “lack of speciation” has been pointed out in the seminal work of Holland back in 1975. It is usually labeled with the oxymoron “premature convergence” to stress the tendency of an algorithm to convergence toward a point where it was not supposed to converge to in the first place. The research activity would tackle “diversity promotion”, that is either “increasing” or “preserving” diversity in an EC population, both from a practical and theoretical point of view. It will also include the related problems of defining and measuring diversity.
Rsearch objectives and methods: This research started about a decade ago with “A novel methodology for diversity preservation in evolutionary algorithms” (G. Squillero, A. Tonda; GECCO Companion; 2008) and is more active every passing year. The proposer organized workshops on diversity promotion in 2016 and 2017; moreover, he presented tutorials on the subject at IEEE CEC 2014, PPSN 2016, and ACM GECCO 2018. This year, together with Dirk Sudholt, a scholar devoted to theoretical aspects of EAs, the research incorporated rigorous runtime analyses. A joint tutorial will be presented at ACM GECCO 2020 (“Theory and Practice of Population Diversity in Evolutionary Computation”).

The primary research objective will be to choose a possible definition of diversity, and to analyze and develop well-known, highly-effective, general-purpose methodologies able to promote it. That is, methodologies like the “island-model”, that are not linked to a specific implementation nor to a specific paradigm but are able to modify the whole evolutionary process. Then the study would examine why and how the divergence of character works in nature, and then find analogies in the artificial environment. Such ideas would be first tested on reduced environment and problems used by scholars. Later, they will be generalized to a wide scenario, broadening their applicability for practitioners. Rigorous runtime analyses would be used to assess the proposed methodologies.

As “premature convergence” is probably the single most impairing problem in the industrial application of EC, any methodology able to ease it would have a tremendous impact. To this end, the proposed line of research is generic and deliberately un-focused, not to limit the applicability of the solutions. However, the research will explicitly consider domains where the proposer has some experience. Namely:
- CAD Applications, mostly related to the generation of Turing-complete assembly programs for test and validation of microprocessors.
- Evolutionary Machine Learning, that is mostly EC techniques used to complement traditional ML approaches.
- Computational Intelligence and games
Outline of work plan: The first phase of the project shall consist of an extensive experimental study of existing diversity preservation methods across various global optimization problems. The MicroGP, a general-purpose EA developed in house, will be used to study the influence of various methodologies and modifications on the population dynamics. Solutions that do not require the analysis of the internal structure of the individual (e.g., Cellular EAs, Deterministic Crowding, Hierarchical Fair Competition, Island Models, and Segregation) shall be considered. This study should allow the development of a, possibly new, effective methodology, able to generalize and coalesce most of the cited techniques.

During the first year, the candidate will take a course in Artificial Intelligence, and all Ph.D. courses of the educational path on Data Science. Additionally, the candidate is required to learn the Python language.

Starting from the second year, the research activity shall include Turing-complete program generation. The candidate will move to MicroGP v4, the new, Python version of the toolkit under active development. That would also ease the comparison with existing state-of-the-art toolkits, such as inspired and deap. The candidate will try to replicate the work of the first year on much more difficult genotype-level methodologies, such as Clearing, Diversifiers, Fitness Sharing, Restricted Tournament Selection, Sequential Niching, Standard Crowding, Tarpeian Method, and Two-level Diversity Selection.

The theoretical aspects of the topic will be better analyzed together with Dr. Dirk Sudholt, from the University of Sheffield. The activity will be likely carried out through the COST project ImAppNIO.

At some point, probably toward the end of the second year, the new methodologies will be integrated into the Grammatical Evolution framework developed at the Machine Learning Lab of University of Trieste – GE allows a sharp distinction between phenotype, genotype and fitness, creating an unprecedented test bench, and G. Squillero is already collaborating with prof. E. Medvet on these topics and jointly published the journal paper “Multi-level diversity promotion strategies for Grammar-guided Genetic Programming” (Applied Soft Computing, 2019).

A remarkable goal of this research would be to definitely link phenotype-level methodologies to genotype measures.
Expected target publications: Journals with impact factors
- ASOC - Applied Soft Computing
- ECJ - Evolutionary Computation Journal
- GPem - Genetic Programming and Evolvable Machines
- Informatics and Computer Science Intelligent Systems Applications
- IS - Information Sciences
- NC - Natural Computing
- TCIAIG - IEEE Transactions on Computational Intelligence and AI in Games
- TEC - IEEE Transactions on Evolutionary Computation
Top conferences
- ACM GECCO - Genetic and Evolutionary Computation Conference
- IEEE CEC/WCCI - World Congress on Computational Intelligence
- PPSN - Parallel Problem Solving From Nature
Required skills and competences:
Current funded projects of the proposer related to the proposal: The proposer is a member of the Management Committee of COST Action CA15140: Improving Applicability of Nature-Inspired Optimisation by Joining Theory and Practice
Possibly involved industries/companies:The CAD Group has a long record of successful applications of evolutionary algorithms in several different domains. For instance, the on-going collaboration with STMicroelectronics on test and validation of programmable devices, does exploit evolutionary algorithms and would definitely benefit from the research. Squillero is also in contact with industries that actively consider exploiting evolutionary machine-learning for enhancing their biological models, for instance, KRD (Czech Republic), Teregroup (Italy), and BioVal Process (France).

Title: Computational Intelligence for Computer Aided Design
Proposer: Giovanni Squillero
Group website: http://www.cad.polito.it/
Summary of the proposal: Computational Intelligence (CI) is playing an increasingly important role in many industrial sectors. Techniques ascribable to this large framework have long been used in the Computer Aided Design field: probabilistic methods for the analysis of failures or the classification of processes; evolutionary algorithms for the generation of tests; bio-inspired techniques, such as simulated annealing, for the optimization of parameters or the definition of surrogate models. The recent fortune of the term “Machine Learning” renewed the interests in many automatic processes; moreover, the publicized successes of (deep) neural networks smoothed down the bias against other non-explicable black-box approaches, such as Evolutionary Algorithms, or the use of complex kernels in linear models. The proposal would focus on the use and development of “intelligent” algorithms specifically tweaked on the need and peculiarities of CAD industries.
Rsearch objectives and methods: The goal of the research is twofold: from an academic point of view, tweaking existing methodologies, as well as developing new ones, specifically able to tackle CAD problems; from an industrial point of view, creating a highly-qualified expert able to bring the scientific know-how into a company, while being also able to understand the practical needs, such as how data are selected and possibly collected. The need to team the experts from industry with more mathematically-minded researchers is apparent: frequently a great knowledge of the practicalities is not accompanied by the adequate understanding of the statistical models used for analysis and predictions.

In more details, the research will almost certainly include bio-inspired techniques for generating, optimizing, minimizing test programs; statistical methods for analyzing and predicting the outcome of industrial processes (e.g., predicting the maximum operating frequency of a programmable unit based on the frequencies measured by some ring oscillators; detecting dangerous elements in a circuit; predicting catastrophic events). The activity is also like to exploit (deep) neural-networks, however developing novel, creative results in this area is not a priority. On the contrary, the research shall face problem related to dimensionality reduction, feature extraction and prototypes identification/creation.
Outline of work plan: The research would start by analyzing a current practical need, namely: “predictive maintenance”. A significant amount of data is currently collected by many industries, although in a rather disorganized way. The student would start by analyzing the practical problems of data collection, storage and transmission, while, at the same time, practicing with the principles of data profiling, classification and regression (all topics that are currently considered part of “machine learning”). The analysis of sequences in order to predict the final event, or rather identify a trigger, is an open research topic, with implication far beyond CAD. Unfortunately, unlikely popular ML scenarios, the availability of data is a significant limitation, a situation sometimes labeled “small data”. The research would need to consider techniques less able to process large amount of information, but perhaps more “intelligent”, and to use all problem-specific knowledge available. The thesis could then proceed by tackling problems related to “dimensionality reduction”, useful to limit the number of input data of the model, and “feature selection”, essential when each single feature is the result of a costly measurement. At the same time, the research is likely to help the introduction of more advanced optimization techniques in everyday tasks.
Expected target publications: Top journals with impact factors
- ASOC – Applied Soft Computing
- TEC – IEEE Transactions on Evolutionary Computationv - TC – IEEE Transactions on Computers
Top conferences
- ITC – International Test Conference
- DATE – Design, Automation and Test in Europe Conference
- GECCO – Genetic and Evolutionary Computation Conference
- CEC/WCCI – World Congress on Computational Intelligence
- PPSN - Parallel Problem Solving From Nature
Required skills and competences:
Current funded projects of the proposer related to the proposal: - The proposer is a member of the Management Committee of COST Action CA15140: “Improving Applicability of Nature-Inspired Optimisation by Joining Theory and Practice”.
- The proposer is collaborating with Infineon on the subjects listed in the proposal. Contract: “Machine Learning techniques for the prediction of failures based on in-situ sensors values”.
- The proposer collaborated with SPEA under the umbrella contract “Colibri”. Such contract is likely to be renewed on precisely the topics listed in the proposal.
Possibly involved industries/companies:The CAD Group has a long record of successful applications of intelligent systems in several different domains. For the specific activities, the list of possibly involved companies include: SPEA, Infineon, ST Microelectronics, Comau (through the Ph.D. student Eliana Giovannitti)

Title: Visual Object Detection across Domains
Proposer: Barbara Caputo
Group website: https://scholar.google.com/citations?user=mHbdIAwAAAAJ
Summary of the proposal: Despite impressive progress in object detection over the last years, it is still an open challenge to reliably detect objects across visual domains. Although the topic has attracted attention recently, current approaches all rely on the ability to access a sizable amount of target data for use at training time. This is a heavy assumption, as often is not possible to anticipate in which domain a detector is going to be used, not to access it in advance for data acquisition. The goal of this project is to bring object detector truly close to the real world by leveraging over learning from auxiliary tasks in order to perform adaptation at test time, from a single target image. The work will develop methodological contribution in this direction, tackling at the same time issues related to real time deployment and relation to other important open challenges such as domain generalization in object detection.
Rsearch objectives and methods: Localizing and recognizing object instances in an image is a fundamental visual task with applications that range from surveillance to medical imaging and industrial manufacturing. Modern object detectors leverage on powerful deep learning architectures with constantly growing performance in accuracy and real time capabilities [1]. An important ingredient of this success is the availability of large amounts of clean labeled data which is however particularly costly for detection since the annotation time for bounding boxes and/or masks is significantly higher with respect to global image labeling [2]. Even when model learning is safely done on reliable data, robustness and generalization at deployment time remains extremely challenging [3]. Indeed, as for other visual recognition tasks, detection algorithms suffer when tested out of the trained domain. Besides variations in illumination, background, object appearance and pose, as well as image quality, specific difficulties also arise from the lack of control on the set of object classes and their co-occurrences in different contexts. Further extreme challenges appear when closing large gaps as from artistic products to real world pictures.
These issues have caught the attention of the community only recently and some noteworthy efforts have been done to cope with it. The results of this growing literature are still scattered across different datasets and there is no general consensus on the most promising baseline architecture or strategy. Moreover, some research questions related to the usually adopted unsupervised domain adaptation (DA) setting remain open. All DA methods operate supposing access to unlabeled target data at training time. Of course the cardinality of the target is crucial: a larger set means more data from which estimating the target distribution and possibly better understand its differences from the known source to close the gap. With the data-hungry detection task this aspect is particularly relevant. However, in many practical applications is not possible to anticipate in which domain a detector is going to be used.
Even when this is feasible, really accessing the target domain in advance for data acquisition may be plainly impossible.
In real-world conditions, even the control on the unique nature of the target may be challenging: samples can be the results of several parallel unrelated acquisitions, and collected under multiple unpredictable scenarios with not smoothly changing conditions.

All in all, the only assumption one can truly safely make is that of having a single target sample at test time for which we want a reliable prediction output. This research proposal starts from this and aims at pushing the limits of cross domain object detection by leveraging over self-supervised learning to perform cross domain adaptation on test data [4]. The project will investigate how to use effectively auxiliary tasks at test time and what auxiliary tasks are more suitable for various settings, how to make adaptation at test time also usable in real time settings, and up to which extent this approach is extendable from a single source domain to multiple source domains, exploring the whole spectrum of learning across different visual domains [5], up to domain generalization [6].

Results will be assessed on publicly available datasets, through participation to international challenges and in the realistic settings of automatic tagging of images collected from social media.

[1] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In ICCV, 2019.

[2] Dim P. Papadopoulos, Jasper R. R. Uijlings, Frank Keller, and Vittorio Ferrari. Extreme clicking for efficient object annotation. In ICCV, 2017.

[3] Haichao Zhang and Jianyu Wang. Towards adversarially robust object detection. In ICCV, 2019.

[4] Fabio M. Carlucci, Antonio D’Innocente, Silvia Bucci, Barbara Caputo, and Tatiana Tommasi. Domain generalization by solving jigsaw puzzles. In CVPR, 2019

[5] Massimilano Mancini, Hakan Karaoguz, Elisa Ricci, Patric Jensfelt, and Barbara Caputo. Kitting in the wild through online domain adaptation. In IROS, 2018

[6] Antonio D’Innocente, Silvia Bucci, Barbara Caputo, and Ta-tiana Tommasi. Learning to generalize one sample at a timewith self-supervision.arXiv preprint 1910.03915, 2019
Outline of work plan: M1-M6: Definition of the experimental testbed, literature review and implementation of existing baselines. Implementation of existing self-supervised auxiliary tasks and study of their suitability for the existing baselines.

M6-M12: Implementation of single source domain object detector with auxiliary task adaptation, test on publicly available databases and comparison with the state of the art. Writing of scientific report on findings of Y1.

M13-M24: Extensions of the deep architecture obtained in Y1 to sparse architectures for the auxiliary task branch, so to aim for real-time adaptation. Assessment of work on the established benchmarks. Writing of scientific report on findings of Y2.

M25- M36. Extensions of the deep architecture obtained in Y2 to multi-source and domain generalization scenarios. Assessment of work on the established benchmarks. Writing of scientific report on overall findings of the project. Writing of PhD thesis.
Expected target publications: It is expected that the scientific results of this thesis will be reported in the top conference in the field of computer vision (IEEE CVPR, IEEE ICCV, ECCV) and machine learning (ICML, ICRL, NeurIPS). At least one journal publication is expected on one of the following international journals: IEEE PAMI, IJCV, CVIU, JMLR.
Required skills and competences:
Current funded projects of the proposer related to the proposal: VIDESEC (CINI)
Possibly involved industries/companies:

Title: Context and Emotion Aware Embodied Conversational Agents
Proposer: Andrea Bottino, Fabrizio Lamberti, Marco Torchiano
Group website: http://www.polito.it
Summary of the proposal: Recent advances in Machine Learning and Artificial Intelligence have resulted in a growing interest towards using Embodied Conversational Agents (ECAs) in Human Computer Interaction (HCI). ECAs are animated virtual characters capable of simulating a human-like face-to-face conversation using natural language processing (NLP) and multimedia communicative behaviors that include verbal and non-verbal clues. The availability of increasingly powerful and connected sensors allows ECAs to access contextual information and interact autonomously with human beings and the environment. Then, the possibility to leverage on a virtual body and voice for the interaction requires as well researchers to include social-emotional components in the design of ECA behavior. In other words, these agents should have a personality, emotions and intentions that should be expressed with voice, hands, head and body movements. At the same time, they should be able to enhance their social interaction with the user by simulating and triggering empathy during the interaction. Given their capabilities, ECAs have the potential to play an increasingly relevant role in a plethora of applications ranging from educational and training environments, health and medical care, virtual assistants in industry and virtual companions in games. Despite that, the current state of the art requires substantial contributions to researchers in order to make ECAs more effective, simple to design and implement, capable of fully expressing and conveying (believable) emotions, and leverage on the (fine-grained) analysis of human affects to create a strong empathic bond with the end users.
Rsearch objectives and methods: Implementing ECA is actually a cumbersome and complex process, which requires taking into account several different elements (NLP, context sensing, emotion modeling, affective computing, 3D animations), which, in turns, involves specific technological and technical skills. Thus, there is the need to develop a simple framework that (i) allows developers to support the rapid design, prototyping and deployment of ECA in a variety of heterogeneous use cases and (ii) is easily extensible, allowing the introduction novel features that expand the current capabilities of ECAs. This framework will be the first relevant outcome of the work.

Then, the research project aims at tackling several issues that can improve the quality and effectiveness of the end-user experience when they interact with ECAs. The projects aims at addressing the following problems.

A substantial portion of our communication is nonverbal. Every day, we respond to thousands on nonverbal cues and behaviors including postures, facial expression, eye gaze, gestures, and tone of voice. However, some of these clues are currently relatively little-used in this context. Affective AI provides robust methods to detect facial expressions from off-the-shelf cameras and in the wild. However, few work has been done to exploit user postures or specific gestures (e.g., nodding can express agreement and a frown can convey disapproval or disagreement).

The extraction of paralinguistic factors such as tone of voice, loudness, inflection, and pitch can provide information about the actual emotional states of the other peer in the communication. Thus, computational mechanisms capable of extracting these variables from the analysis of the user’s voice are sorely needed. The same paralinguistic factors should be available to modulate the ECA response according to its emotional states. On the contrary, one of the problems with present text-to-speech libraries is that they pronounce everything with the same tone, which makes impossible to communicate feelings through voice. To this end, one of the possibility that will be explored in the research is to exploit style-transfer approaches (similar to the ones implemented in the visual contexts for, e.g., transferring the style of a painting on a photo taken with a digital camera) capable of “transferring” recognizable emotional styles from a set of reference samples to the synthesized utterance.

The general approach in managing voice interaction between the user and the ECA is to define a set of intents, each associated with a training set of utterances for recognition. However, identifying intents and creating a suitable training set is complex and tedious. Thus, a relevant contribution to soften this issue would be the development of automatic or semi-automatic approaches for harvesting, mining and clustering textual corpus related to the specific scenario addressed.

Finally, the project foresees the development of advanced Computer Vision and Machine Learning approaches to allow ECAs exploit visual information for capturing users’ affect and actions and use these pieces of information to proactively respond to users intents and improve the ECA empathic and emotional response. The same visual mechanism will be also used to obtain low-cost captures of human movements and expressions that can be used to realistically anime the ECAs.
Outline of work plan: In the first year, the project will start from a systematic review of the current state of the art of ECAs. The outcome of this works will serve to analyze the current limitations, identify promising approaches and inform the design and development of the solutions aimed at addressing the problems addressed by the research proposal. The candidate will also start designing, developing and assess a framework that allows the rapid prototyping and deployment of ECA in different contexts.
In the following years, the candidate will develop solutions addressing the aforementioned research objectives. These solutions will be aimed at extending the capabilities offered by this framework and, ultimately, the features that ECAs can exploit to improve the quality and effectiveness of the interaction with the end-users. The proposed techniques will be applied to relevant use-cases exploiting VR and AR applications and will be validated by means of user studies involving panels of volunteers.
Expected target publications: International peer-reviewed journals in the fields related to the current proposal, such as: IEEE Transactions on Affective Computing, IEEE Transactions on Visualization and Computer Graphic, IEEE Transactions on Human-Machine Systems, ACM Transactions on Graphics, IEEE Transactions Image Processing, IEEE transactions Pattern Analysis and Machine Intelligence, Pattern Recognition, Pattern Recognition Letter, Computer Vision and Image Understanding. Relevant and well-reputed international conferences, such as: IEEE Face and Gesture Recognition (FG), ACM Intelligent Virtual Agents, IEEE VR, CVPR, Eurographics.
Required skills and competences:
Current funded projects of the proposer related to the proposal: -
Possibly involved industries/companies:-

Title: Virtual character animation from video sequences
Proposer: Fabrizio Lamberti
Group website: http://grains.polito.it/
Summary of the proposal: Creating realistic virtual character animations is a complex task. Performance-based methods based on motion capture allow to register accurate and realistic movements. Recently, sophisticated deep learning models have been proposed to train character controllers to reproduce various motor skills from motion capture data. However, acquiring such data requires dedicated hardware and a complex and time-consuming setup. Making computer animation accessible to a wide range of non-skilled users requires simplified, intuitive interfaces and animation tools.

This proposal targets techniques for training virtual character controllers that can reproduce a wide range of actions, which may also involve interaction with external objects, by exploiting a generative adversarial imitation learning framework to learn from large scale monocular video datasets. The candidate will study how to reproduce a single action demonstration, as well as multiple animations corresponding to sequences of actions, such as sitting, running or picking up objects.

The candidate will apply the proposed methods to relevant case studies in CG, VR or AR, e.g. the control of realistic Embodied Conversational Agents (ECAs) and/or Non-Player Characters (NPCs) for serious gaming applications.
Results will be evaluated in terms of quality and believability, user effort and computational requirements.
Rsearch objectives and methods: Videos provide a cost-effective and scalable alternative to motion capture data [1]. Learning from multiple human demonstrations also results in more realistic and believable avatars. Hundreds of hours of video material are uploaded to YouTube on a daily basis, and publicly available datasets include numerous examples of hundreds of different types of actions. Recent papers have demonstrated the feasibility of extracting a 3D controller from a video sequence representing a given motor skill, including everyday motions, sports and acrobatic tricks [2-3]. Most approaches rely on recent advances in 2D and 3D pose estimation, such as OpenPose and the more recent Human Mesh Recovery (HMR) model [4].

While the results are exciting, there are still many open issues in recovering 3D human motion from videos and extending the available techniques to modeling of complex tasks involving scene and multiple character interactions. The results are conditioned by the quality of the 3D pose estimation, which is still limited in the presence of occlusions and self-occlusions. The available methods are limited to short sequences, in the type and amount of actions that can reproduce, and usually do not model interaction with the surrounding scene.

The goal of this project is to develop new techniques for training virtual character controllers that can reproduce a wide range of actions, which may involve interaction with external objects, by exploiting a generative adversarial imitation learning framework to learn from large scale video datasets. The goal is to be able to reproduce a single demonstration, as well as to be able to generate animations associated to a sequence of actions, such as sitting, running or picking up objects. In this way, it will be possible to construct complex animations by providing a script or a list of actions, and the corresponding animations will be merged. By exploiting a generative setting, it will be possible to better capture the intrinsic variability of human motions, and hence produce animations that look more natural and realistic. Techniques that can adapt the quality of the motion to a specific person (style transfer) or to emotional states may be also analyzed.

[1] Peng, X.B. et al, 2019. Sfv: Reinforcement learning of physical skills from videos. ACM Transactions on Graphics (TOG), 37(6), p.178.
[2] Starke, S., Zhang, H., Komura, T. and Saito, J., 2019. Neural state machine for character-scene interactions. ACM Transactions on Graphics (TOG), 38(6), p.209
[3] Kanazawa, A., Zhang, J.Y., Felsen, P. and Malik, J., 2019. Learning 3d human dynamics from video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 5614-5623).
[4] Kanazawa, A., Black, M.J., Jacobs, D.W. and Malik, J., 2018. End-to-end recovery of human shape and pose. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 7122-7131).
Outline of work plan: The candidate will study novel techniques to produce animate virtual characters from monocular video sequences by building on a generative adversarial imitation learning framework. To enhance the feasibility, the problem will be decomposed in two sub-problems.

First, video understanding and interpretation techniques will be developed to recognize temporal sequences of actions from live demonstrations, as well as interactions between with surrounding objects and environments (semantic level). To this aim, state-of-the-art algorithms for constructing parse graphs for image and video analysis, as well as in action recognition and temporal localization, will be leveraged.

Secondly, the PhD candidate will work on reproducing fine-grained realistic animations for specific actions (motion level). This will require to extend semantic level image interpretation with fine-grained 3D mesh and pose recovery, beyond the current state of the art.

Different neural models will be designed, implemented and compared in order to generate animations (3D poses) for individual actions, which will then be blended to achieve a smooth transition between different actions. Starting from simple sequences and single actions, the candidate will progressively move on to more complex scenes that require interaction of multiple characters, as well as with the surrounding 3D scene. This will require to extend existing computer vision and machine learning techniques not only to extract adequate supervisory signals from videos, but also to handle interactions and collisions in the virtual world.

A semi-supervised learning approach will be adopted to leverage existing large-scale annotated datasets for action recognition, combined with available pose estimation techniques (as opposed to manually annotating the 2D poses). Video data may be augmented with motion capture and/or 3D datasets, such as ShapeNet, especially for the generation of additional 3D object representations.

The proposed techniques will be applied to relevant case studies, possibly integrated in AR/VR applications (e.g., for health, training and emergency scenarios). Applications include the personalization of ECAs and/or the control of realistic NPCs for serious games. Results will be assessed in terms of animation quality and compared with the user effort, ease of use and computational resources required.
Expected target publications: The target publications will cover computer vision, computer graphics, and human-computer interaction journal and major conferences. Journals could include:

International Journal of Computer Vision
Computer Vision and Image Understanding
IEEE Transactions on Human-Machine Systems
IEEE Transactions on Visualization and Computer Graphics
ACM Transactions on Graphics
IEEE Transactions on Pattern Analysis and Machine Intelligence
Required skills and competences:
Current funded projects of the proposer related to the proposal: - E2DRIVER (H2020), on the use of VR for training in industrial settings.
- VRRobotLine (research grant from Bell Production Spa in the context of a Regional project), on the use of VR for robotic applications.
- Research grant from SIPAL SpA, on the usage of AR in the engineering domain.
Possibly involved industries/companies:- KUKA Roboter Italia Spa
- Italian Airforce
- WpWeb Srl

Title: Robust machine learning models for high dimensional data interpretation
Proposer: Fabrizio Lamberti
Group website: http://grains.polito.it/
Summary of the proposal: The successes of deep neural networks (DNNs) in the analysis of multi-dimensional data is largely due to their ability in modelling high-dimensional objects (such as pixels in an image) by learning non-linear feature representations, and the availability of simple and scalable supervised learning algorithms. Nonetheless, DNNs are often surprisingly brittle in practice and lack desirable properties including top-down control, transparency and robustness.

There is an increasing interest in complementing representation learning with techniques borrowed from statistical relational reasoning and knowledge representation. Symbolic knowledge representation excels precisely where statistical machine learning is weakest, in terms of transparency, ability to perform inference and reasoning, representation of high-level concepts, integration of prior knowledge and, in general, model of the external worlds.

Recent advances in computationally efficient statistical inference have also significantly expanded the toolbox of probabilistic modeling techniques that can be applied to deep learning models. This has important practical implications in terms of robustness. One of the most profound effects is, for instance, the ability to “know when you don’t know”.

The PhD candidate will study techniques that incorporate deep learning and probabilistic reasoning for the interpretation of multi-dimensional data, such as images.
Rsearch objectives and methods: The overall objective of this research proposal is to explore novel ways to bridge the gap between deep representation learning and symbolic knowledge representation in multi-dimensional data analysis, leveraging recent advances in the field of neural-symbolic integration and probabilistic modeling.

Specifically, this research proposal targets the area of neural-symbolic integration, which can be used to encode symbolic representation techniques, such as fuzzy and/or probabilistic logic, as tensors in a deep neural network. This would allow to extend current DNNs feature learning capabilities in many ways, for instance by imposing a priori information.

Many recent techniques have been proposed to combine deep learning with reasoning, but the applications are still limited compared to their potential. In many of the proposed approaches, either the feature representation or the reasoning layer is frozen and cannot be easily learned, and/or feature extraction and reasoning are disjoint and cannot be trained end-to-end. Probabilistic reasoning is also required to handle uncertainty and exceptions in real-life applications. It allows to quantify uncertainty in the predictions and thus, enable the development of robust machine learning models. Finally, application in large-scale datasets and a variety of tasks beyond classification is still limited.

The candidate will target the application of neural-symbolic integration techniques to solve problems in different image analysis tasks, such as image-level classification, segmentation and object detection. Of particular interest is the possibility to explore the integration of image level-data with other sources of information, for instance the integration of prior (possibly causal) information.

There are a number of theoretical and practical issues to overcome to apply this approach at scale. First, techniques are needed that can achieve both representation learning and relational reasoning, and that can be trained in an end-to-end fashion at scale. Secondly, the above-mentioned goals must be achieved in a probabilistic framework, leveraging recent advances in statistical inference and optimization. During the PhD, new solutions will be studied and developed to address the issues listed above, and finally compared with state-of-the-art deep learning approaches to assess their potential on machine learning benchmarks.

The proposed techniques will also be applied in at least one selected case study, such as medical image analysis and its integration with other clinical data, or problems related to autonomous driving. To this aim, the PhD candidate will benefit from existing collaborations with industrial and clinical partners, as well as within the SmartData@PoliTO research center.
Outline of work plan: Phase 1: the candidate will strengthen core competencies in deep learning / machine learning, statistical relational reasoning, representation learning and statistical inference. He/she will survey the relevant literature to identify and compare existing frameworks and methods for neural-symbolic integration; an evaluation of the most important research gaps is expected as output. The candidate will design neural-symbolic architectures to solve benchmark tasks in computer vision, such as image classification, and compare them with standard learning techniques. In this phase, the candidate will work on established machine learning and computer vision benchmarks, starting from simpler datasets such as CIFAR100 and FashionMNIST and moving on to more complex datasets, to compare the proposed solutions with state-of-the-art convolutional neural networks in terms of accuracy, convergence properties, interpretability and generalizability.

Phase 2: the candidate will further extend the proposed methods by integrating probabilistic modeling techniques with a focus on estimating uncertainty and integrating priors in deep learning models. For instance, the advantages of integrating a priori information from WordNet or other sources to improve the accuracy of classification models will be evaluated. The proposed techniques will be compared with state-of-the-art deep learning techniques in terms of performance, generalizability, training speed, transparency and robustness.

Phase 3: the candidate will apply the proposed techniques to one or more selected case studies to evaluate their applicability in real-life scenarios. In this phase, one case study pertaining to GRAINS research projects will be selected, and the candidate will closely collaborate with other researchers and domain experts to define and train the model, as well as to import existing knowledge bases in the neural-symbolic architecture. For instance, the GRAINS research group is active in the field of computer-aided detection and diagnosis, and screening mammography could be likely a candidate for the case study.
Expected target publications: Publications will cover conferences and journals in machine learning and computer vision. Target journals include: IEEE Transactions on Pattern Analysis and Machine Intelligence, International Journal of Computer Vision, IEEE Transactions on Neural Networks and Learning Systems, Pattern Recognition, Computer Vision and Image Understanding, Journal of Machine Learning Research
Required skills and competences:
Current funded projects of the proposer related to the proposal: - Project AIBIBANK: Bio-Banking for Artificial Intelligence in Oncology (submitted to the call Pi.Te.F. PIATTAFORMA TECNOLOGICA DI FILIERA, in collaboration with DISAT department (ref. Prof. Andrea Pagnani), Città della Salute e della Scienza and IRCCS Candiolo).
- Research contract with FCA-CRF on “Obstacle detection and trajectory forecasting”.
Possibly involved industries/companies:

Title: Urban intelligence
Proposer: Silvia Chiusano
Group website: http://dbdmg.polito.it
Summary of the proposal: Due to the rapid development of cities and the increase in urban population, past decades have seen many urban issues arise, such as traffic congestion, energy shortages and air pollution. At the same time, an increasing volume of a variety of data is generated and collected with the help of new technologies.

Urban intelligence entails the acquisition, integration, and analysis of big and heterogeneous data collections generated by a diversity of sources in urban spaces to profile the different facets and issues of the urban environment.

Digging deep urban data collections can unearth a rich spectrum of knowledge valuable to characterize citizen behaviours, identify weaknesses and strengths of the services provided in the urban area as well as improve the quality of these services or even devise new ones.

However, data analytics on these data collections is still a daunting task, because they are generally too big and heterogeneous to be processed through data analysis techniques currently available. Consequently, from a data science perspective, data emerging from today's urban environment give rise to a lot of challenges that constitute a new inter-disciplinary field of research.
Rsearch objectives and methods: The PhD student will work on the study, design and development of proper data models and novel solutions and for the acquisition, integration, storage, management and analysis of big volumes of heterogeneous urban data.

The research activity involves multidisciplinary knowledge and skills including database, machine learning techniques, and advanced programming. Different case studies in urban scenarios such as urban mobility, citizen-centric contexts, and healthy city will be considered to conduct the research activity. The objectives of the research activity consist in identifying the peculiar characteristics and challenges of each considered application domain and devise novel solutions for the management and analysis of urban data for each domain. More urban scenarios will be considered with the aim of exploring the different facets of urban data and evaluating how the proposed solutions perform on different data collections.

More in detail, the following challenges in the context of data analytics will be addressed during the PhD, such as:

- Suitable data fusion techniques and data representation paradigms should be devised to integrate the heterogeneous collected data into a unified representation describing all facets of the targeted domain. For example, urban data are usually acquired by means of different sensor networks deployed in the city. Since these data are often collected with different spatial and temporal granularities, suitable data fusion techniques should be devised to support the data integration phase, and provide a spatio-temporal alignment of collected data.

- Adoption of proper data models. The storage of heterogeneous urban data collections requires the use of alternative data representations to the relational model such as NoSQL databases (e.g., MongoDB).

- Design and development of algorithms for big data analytics. Urban data is usually charaterized by spatio-temporal coordinates describing when and where data has been acquired. Spatio-temporal data has unique properties, consisting of spatial distance, spatial hierarchy, temporal smoothness, period and trend, which entails the design of suitable data analytics methods. Moreover, huge volume of data demands the definition of novel data analytics strategies also exploiting recent analysis paradigms and cloud based platforms as Hadoop and Spark.

- Proper strategies will be also devised for data and knowledge visualization.
Outline of work plan: 1st Year. The PhD student will review the recent literature on urban computing to identify the up-to-date research directions and the most relevant open issues in the urban scenario. Based on the outcome of this preliminary explorative analysis, an application domain, such as urban mobility or urban air pollution, will be selected as a first reference case study. This domain will be investigated to identify the most relevant data analysis perspectives for gaining useful insights and to assess of main data analysis issues. The student will perform an exploratory evaluation of state-of-the-art technologies and methods on the considered domain, and will present a preliminary proposal for the optimization techniques of these approaches.

2nd and 3rd Year. Based on the results of the 1st year activity, the PhD student will design and develop a suitable data analysis framework including innovative analytics solutions to efficiently extract useful knowledge in the considered domain, aimed at overcoming weaknesses of state-of-the-art methods.

Moreover, during the 2nd and 3rd year, the student will progressively consider a larger spectrum of application domains in the urban scenario. The student will evalute if and how his/her proposed solutions can be applied to the new considered domains as well as he/she will propose novel analytics solutions.

During the PhD, the student will have the opportunity to cooperate in the development of solutions applied to the research projects on smart cities. The student will also complete his/her background by attending relevant courses. The student will participate to conferences presenting the results of his/her research activity.
Expected target publications: Any of the following journals
IEEE TKDE (Trans. on Knowledge and Data Engineering)
ACM TIST (Trans. on Intelligent Systems and Technology)
Journal of Big Data (Springer)
Expert Systems With Applications (Elsevier)
Information Sciences (Elsevier)

IEEE/ACM International Conferences
Required skills and competences:
Current funded projects of the proposer related to the proposal: CANP (La Casa nel Parco): Regione Piemonte, Piattaforma Tecnologica “Salute e Benessere”

S[m2]ART (Smart Metro Quadro): Bando Smart City and Communities; Ente finanziatore: MIUR (Ministero dell’istruzione, Università e Ricerca)

MIE (Mobilità Intelligente Ecosostenibile): Bando Cluster Tecnologico Nazionale “Tecnologie per le Smart Communities”; Ente Finanziatore: MIUR (Ministero dell’istruzione, Università e Ricerca)
Possibly involved industries/companies:

Title: Energy-Efficient Conditional Inference of Sequential Data
Proposer: Enrico Macii, Massimo Poncino
Group website: http://eda.polito.it
Summary of the proposal: Sequential Data is any kind of data where the relative order matters, and time series can be considered as a subset thereof, where the order is determined by time stamps. Many applications produce data sets that fit well the sequential category: speech, text, videos, or any temporal series. Analysis of sequential can be done with classical Machine Learning (ML) algorithms (e.g., decisions trees, SVMs, state-space models) but deep neural networks are now widely used and achieving state-of-the-art results. In the latter case, the computational cost can be very high, as these networks are usually iterative (recurrent neural networks, RNN) and therefore the option of moving the computation to the edge to achieve energy efficiency and more deterministic latency becomes even more challenging than for traditional feed-forward DNNs. Moreover, the vast majority of existing efforts in moving machine/deep learning algorithms to the edge are focused on classification of individual data items (e.g., images) and not on sequential models.

In this proposal, we will explore one of the most promising ways to achieve this result, that is, to resort to the so-called conditional inference methods (also known as staged inference or hierarchical inference). These solutions are based on appropriately combining multiple ML models of different computational complexity (and accuracy), under the assumption that data are not all equally challenging: the simplest (and less demanding) algorithm are executed most of the times on edge nodes, which allows to respect the latency and energy constraints of the application. Whenever the output of these algorithms is not satisfying (e.g. a classification confidence is too low), more accurate and demanding algorithms are invoked. These can still run on edge nodes (impacting minimally on latency and energy due to their rare invocations) or be executed in the cloud.
Rsearch objectives and methods: The final goal of the thesis will be the development of a framework for implementing machine learning tasks on resource-constrained edge nodes, using conditional inference concepts to be applied on sequential data. While there are some preliminary works on this topic in the literature, the objective of this research will be to apply this concept systematically at all levels of the computing stack.

The framework should be flexible, so that it can be applied to different application domains, but the main case studies will be related to Industry 4.0 applications. Use cases that we are already experienced include the analysis of time series of sensor data for predictive maintenance applications, but also sequences of images relative to defective objects or video frames for moving objects on a conveyor belt.

Conditional inference techniques will be developed at different levels of the computing stack. The following is a (non-comprehensive) list of possible approaches, taken from the existing ones in the literature. More advanced solutions will be developed over the course of the research.

At the algorithm level, the use of models of different size and accuracy is the straight-forward solution to implement conditional inference. These multiple models can be of homogeneous type (e.g. multiple neural networks with different number of layers and features) or heterogeneous (e.g. a simple but efficient binary tree combined with a more accurate but more complex neural network). Techniques from ensemble learning (e.g. the super-learner) will be used for combining the outputs of these multiple models and to select which of them to use for a given set of input data. Alternatively, selective refinement techniques can also be used, where the high-complexity model is only executed on (particularly challenging) portions of the input data

At the level of data representation, conditional learning can be achieved by using different data formats in different “stages”, such as low-precision integer data for the low-complexity stages and floating-point for the high-complexity ones. Extreme data-representation reduction techniques could also be used (e.g. binary/ternary neural networks).

Finally, at the hardware level, staged inference can be implemented by selectively executing an inference either on a general purpose embedded CPU or on a dedicated hardware accelerator (embedded GPU or NPU).

Clearly, integration among these different abstraction levels will be a fundamental element of the developed framework. For instance, considerations on the computational requirements of different types of ML algorithms will depend on the target hardware and be influenced by the chosen data representation. As an example, Temporal Convolutional Networks (TCNs) can be used as low-complexity alternatives to Recurrent Neural Networks (RNNs) due to the more hardware-friendly and parallelizable operations involved in their execution.

The candidate will mostly work in C/C++ and Pytho and should have a basic understanding of machine learning models and of computer architectures.
Outline of work plan: PHASE I (months 1-9):
- Study of the state of the art in energy-efficient machine learning
- Study of existing conditional learning technique and of methods for sequential data
- Analysis of the characteristics of the target Industry 4.0 use cases (predictive maintenance, automated quality control, etc.)

PHASE II (months 9-18):
- Design and implementation of individual conditional learning techniques offline (i.e, on PCs)
- Deployment on mobile and IoT devices.

PHASE III (months 18-36):
- Development of a comprehensive framework to implement conditional learning, by combining previously designed techniques
- Assessment of the benefits deriving from the developed framework on a set of real industrial use cases.
Expected target publications: - IEEE Transactions on Computers
- IEEE Transactions on CAD
- IEEE Design and Test of Computers
- ACM Transactions on Embedded Computing Systems
- IEEE Transactions of Design Automation of Electronic Systems
- IEEE Pervasive Computing
- IEEE Journal on Internet of Things
Required skills and competences:
Current funded projects of the proposer related to the proposal: - Regional Project "DigitalTwin" (Jan 2020)
- ECSEL Project “MADEin4” (Apr 2019)
Possibly involved industries/companies:Reply, STMicroelectronics

Title: Development of Agent Based Models for smart energy policies in energy communities
Proposer: Enrico Macii, Lorenzo Bottaccioli
Group website: http://eda.polito.it
Summary of the proposal: A smart citizen-centric energy system is at the center of the energy transition in Europe and worldwide. Local energy communities will enable citizens to participate collectively and actively in local energy markets. New energy polices such as Demand Response and Demand Side Management will be adopted to manage the grid stability and to reduce both energy waste and bill. Moreover, new digital tools (e.g., smart energy contracts) will be used to manage financial transactions connected to the exchange of energy among different prosumers and with the grid. On the one side, digital and energy technology combined together will provide a framework for a more intelligent and sustainable final use of energy in buildings and cities. On the other side, citizens will need to understand how to interact with smart energy systems and local energy markets. Given this emerging panorama, it will become even more important the understanding on the dynamics of energy technology diffusion among potential adopters, and the impact that regulation and policy could have on diffusion patterns and penetration levels at the single citizen/household level. Hence, the candidate will develop and ABM co-simulation platform that will integrate heterogenous simulator in order to simulate and understand such phenomenon. Moreover, the candidate will investigate and develop the required agents intelligence by exploiting novel machine learning algorithms and theory from social and psychological modelling.
Rsearch objectives and methods: The diffusion of distributed (renewable) energy sources poses new challenges in the underlying energy infrastructure, e.g., distribution and transmission networks and/or within micro (private) electric grids. The optimal, efficient and safe management and dispatch of electricity flows among different actors (i.e., prosumers) is key to support the diffusion of distributed energy sources paradigm. The goal of the project is to develop a distributed co-simulation agent-based modeling framework to describe the complexity of a smart multi energy system. State of art solution of ABM in smart-energy lack in integrating together the operational and planning phase of future smart energy systems. Moreover, there is a lack in simulating the diffusion and the growth of Energy communities and or smart energy strategies such as demand response or demand side management.

Hence, the research program will focus on the development of the algorithms that will characterize the behaviors of the different agents to describe:
1) the final customer/prosumer beliefs desire and intention and opinions.
2) the local energy community coordinator that has to manage and optimize the energy consumption/production
3) the local energy market where prosumers can trade their energy and or flexibility
4) the local system operator that has to provide the grid reliability
5) the local energy resource/storage that are present in the community

All the software entities will be coupled with external simulators of grid, market, energy sources in a plug and play fashion.

The final outcomes of this research will be a scalable agent-based platform that can be exploited for:
- Planning the evolution of future smart multi energy system by taking in to account the operational phase
- Evaluating the effect of different policies and related customer satisfaction
- Evaluating the diffusion of technologies and/or energy policies under different regulatory scenarios
- Evaluating new business model for energy communities and aggregators
Outline of work plan: 1st year. The candidate will study state-of the-art solution of existing agent-based frameworks in order to identify the best available solution for large scale smart energy system simulation in distributed environments. Furthermore, the candidate will review the state of the art in prosumers/aggregators/market modelling in order to identify the challenges and identify possible innovations. Moreover, the candidate will focus on the review of scheduling optimization techniques with a particular focus on the integration of prosumers preferences (e.g. reinforcement learning). Finally, the candidate will design the overall platform starting for the requirements identification and it will start the development of the first agents intelligent.

2nd year. The candidate will finish the implementation of the agents' intelligence. Furthermore, the candidate will start to integrate agents and simulators together in order to crate the first beta version of the platform with a simple case study.

3rd year. The candidate will ultimate the overall platform and test it in different case study and scenarios in order to demonstrate all the capabilities of the implemented solution.
Expected target publications: - IEEE Transaction Smart Grid
- IEEE Transactions on Industrial Informatics
- IEEE Transactions on Emerging Topics on Computing
- IEEE Transactions on Evolutionary Computation
- IEEE Systems Journal
- Environmental modelling and Software
- ACM e-Energy
Required skills and competences:
Current funded projects of the proposer related to the proposal:
Possibly involved industries/companies:

Title: Anomaly detection and knowledge extraction in household energy consumption
Proposer: Enrico Macii, Edoardo Patti
Group website: http://eda.polito.it
Summary of the proposal: The recent technological advances in pattern recognition, machine learning and data analytics disrupted several fields of modern society. Among them, the energy sector is one of the most promising, in which new technologies would benefit from the insights deriving from the application of such techniques to household energy consumptions.

This research proposal concerns the development of (semi-) unsupervised models, focusing on household appliances load profiles. While in use, each appliance leaves a digital footprint behind, which embeds valuable information, such as programs, energy class labels and potential failures of the system. Accessing this precious knowledge would lead to enable a number of new services and business models that would revolutionize the future energy marketplace, e.g. energy consumption forecast, Demand Response, Demand Side Management and State Estimation.

On the one hand, such models can simultaneously target two goals: anomaly detection and knowledge extraction. This information is crucial to define the energy consumption patterns of customers improving the quality of further algorithms for smart grid management. Moreover, such kind of information could enhance the awareness of the final customers about their consumptions, entailing fewer waste of energy.

On the other hand, such models can generate realistic synthetic load profiles for different appliances based on the learnt knowledge (e.g. use, energy class etc.).
Rsearch objectives and methods: The objectives of the Ph.D. plan are the following:
1. Developing the competences to design (semi-) unsupervised models to derive useful information from household load consumption
2. Providing a scalable solution to be applied to the common household appliances either energivorous or not.
3. Presenting a scalable and easy-to-apply methodology to generate realistic and synthetic load profiles of heterogeneous appliances in terms of type, use and energy class.
4. Providing final results with a limited (i.e., acceptable) error with respect to baseline data.

The final outcome of these algorithms can be exploited to
- accurately forecast load consumption profiles of users,
- evaluate new control policies for smart grid management (e.g. Demand Response, Demand Side Managements and State Estimation),
- to feed simulation engines to realistically mimic the behavior of energy distribution networks,
- to realistically evaluate new business models and applications even considering the spread of new and more efficient appliances,
- to provide personalized tips and suggestions to customers even based on the anomaly behaviors of their appliances.

The aforementioned research activities will focus on two main area of application in computer science, in particular to pattern recognition, machine learning and data analytics.
Outline of work plan: 1st year. The candidate will study state-of the-art techniques to i) derive information and meta-data in household energy consumptions, e.g. programs, energy class labels of appliances and ii) detect potential failures and anomalies. At the beginning, various techniques should be considered, to classify and cluster different load energy consumption patterns of household appliances in order to prove the effectiveness of the basic approach. In this phase, the most energivorous appliances will be considered (e.g. washing machines, dishwashers and fridges).

2nd year. Based on the outcomes of the first year, the developed algorithms will be extended i) to cover more type of appliances, even more energy efficient, and ii) to discern across different possible anomalies for each appliance taken under consideration.

3rd year. The methodology and the algorithms developed in the previous years will be validated to prove their robustness and scalability in being applied on a large set of different possible appliance load profiles.
Expected target publications: - IEEE Transactions on Smart Grids
- IEEE Transactions on Industrial Informatics
- IEEE Transactions on Emerging Topics on Computing
- IEEE Transactions on Computers
- IEEE Systems Journal
- Pattern Recognition
- Expert systems with applications
Required skills and competences:
Current funded projects of the proposer related to the proposal: Arrowhead-tools
Possibly involved industries/companies:Midori

Title: Deep-Learning on the Edge of the IoT: optimization methods, tools and circuits for ultra-scalable platforms
Proposer: Enrico Macii, Andrea Calimera
Group website: http://eda.polito.it
Summary of the proposal: The recent breakthrough on Deep Learning (DL) represents a great opportunity for the evolution of the IoT. Together, IoT and DL enable the implementation of intelligent infrastructures that can sense the environment, predict upcoming events or future trends or users’ requests, and thus take decisions that optimize the use of global resources.

In most of today’s implementations, inference models based on DL algorithms are processed in the cloud, namely, on high-performance computing platforms placed in datacenters and server-farms. Unfortunately, such a cloud-centric vision of the IoT has proven quite inefficient in several ways. A more sustainable growth of smart services will indeed require shifting the processing of DL into the sensor nodes. This concept of “data-analytics on the edge” is an emerging paradigm thought to decentralize and distribute the computational effort and thus to ensure faster and more certain response time, less energy waste, and improved privacy of data.

The challenge faced by this research proposal is to fit the complexity of DL algorithms into tiny processing nodes with low energy budgets and tight resource constraints. More specifically, the proposal deals with the development of a new class of design strategies, optimizations tools and methods that will enable the design and deployment of DL on ultra-low power portable devices.
Rsearch objectives and methods: The broad objective of the project is to bring complex data-analytics, DL models in particular, on the mobile edge of the IoT. This will be achieved by enabling the processing of Deep Neural Networks (DNNs) on lightweight computer architectures that can be deployed on low-power/low-cost end-nodes. From a technical viewpoint, this encompasses the development of new hw/sw co-design strategies and optimization methods which span the whole design hierarchy. The problem will be therefore attacked by two opposite directions: from the top, at the software level, with the development of new learning strategies that can shrink the complexity of DL models; from the bottom, at the hardware level, with the design of digital architectures able to reach an energy efficiency of few pJ/operation. This translates in two main technical objectives:
1. Development of new training techniques targeting multiple constraints, including performance, energy and memory footprint. Such design constraints are strongly interrelated and represent conflicting objectives in the optimization space. The techniques developed will represent the new smart tools by which it will be possible to find the best trade-off between functional and extra-functional metrics.
2. Development of tools and optimization strategies for the design of a new class of circuits specialized to run neural tasks, like dedicated ultra-low power hardware architectures that accelerate (in terms of throughput/power) the processing of DL models leveraging dedicated primitives and new memory management strategies.

The main activities of the candidate will thereby cover the following topics:
1. Multi-objective training, where the optimization objective includes not only the prediction accuracy of the inference engine but also extra-functional metrics.
2. CAD tools for neural architectures, a set of hw/sw design and analysis tools for specialized neural accelerators that enable to meet the energy/throughput requirements of IoT applications.
3. Design of hardware accelerators, custom digital components to be integrated with existing general-purpose cores in order to accelerate specific neural tasks.
Outline of work plan: 1. State of the art (6 - 9 months): Analysis and study of the state of the art in the field of machine-learning, DL models and training methods, DL optimization strategies on lightweight CPUs.
2. Development of multi-objective learning algorithms (duration: 8 - 12 months): such algorithms will be formalized as a max-min problem, namely, minimize model representations while maximizing the classification accuracy. The obtained DL model will then assume a computational-friendly representation, thus, to allow an effortless evaluation through lightweight operations. At the same time, it will also be compliant with strict memory requirements of IoT devices, thus enabling an easy migration of complex models to resource-constrained environments.
3. Optimization Framework (duration: 12 - 18 months): automation of design flow and optimization algorithms to achieve low-power and high-throughput processing of DL models; the resulting framework addresses power optimization at different levels of abstraction, from the software level to the circuit level.
4. Design of digital architecture for DL (duration: 12 - 18 months): implementation of a specialized accelerators that can host and run complex DNN models with few resources; the new architecture will be conceived considering power/energy consumptions and throughput as the main constraints.
5. Integration and Validation (duration: 6 - 9 months): the implemented architecture and optimization tools will be tested on the use-cases; quality assessment will be done considering accuracy of data-analysis and energy consumption as the main metrics.

The above activities should not to be considered as stand-alone tasks; they may overlap, indeed, in order to achieve a more efficient integration and a higher quality of results.
Expected target publications: Journals and Conferences/Workshops whose topics include but are not limited to Design Automation, Artificial Intelligence, Machine Learning, Computer Architectures, Communication and Networking, VLSI circuits. Some example below:
- ACM TRANSACTIONS ON DESIGN AUTOMATION OF ELECTRONIC SYSTEMS
- IEEE TRANSACTION ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS
- IEEE TRANSACTIONS ON NEURAL NETWORKS
- IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS
- IEEE TRANSACTIONS ON COMPUTERS
- IEEE TRANSACTIONS ON VLSI
- IEEE TRANSACTIONS ON COMMUNICATION
- IEEE/ACM TRANSACTIONS ON NETWORKING
Required skills and competences:
Current funded projects of the proposer related to the proposal: There are two on-going EU projects whose research goals do match with this proposal:
- R3-PowerUP: 300nm Pilot Line for Smart Power and Power Discretes (funded by the European Commission through H2020 - ECSEL framework)
- MADEin4: Metrology Advances for Digitized ECS industry 4.0 (funded by the European Commission through H2020 - ECSEL framework)
In both the projects, there are research activities which include the implementation and deployment of predictive models into low-power/low-cost micro-computers.
Possibly involved industries/companies:STMicroelectronics

Title: Development of infrastructures, simulation models and data analytics for manufacturing process optimization
Proposer: Enrico Macii, Elisa Ficarra
Group website: http://eda.polito.it
Summary of the proposal: European industry is facing substantial economic challenges due to an ever increasing rate of growth in technological development, a growing globalization of markets and competition with productive realities regulated by different costs and labor policies. These issues are driving towards the development of industrial technologies to reduce the workforce, use resources efficiently, shorten manufacturing lead time and throughput time. This phase requires the development of technologies for Cyber-Physical Systems, IoT, cloud computing and automation capable of producing and collecting extended data in Volume, Speed and Variety, globally defined as Industrial Internet of Things (IIoT). Furthermore, there is the need to develop techniques for processing such data to extract information useful for making the industrial manufacturing process automated, controlled and efficient.

The purpose of this doctoral proposal is therefore the integration of these needs through the development of infrastructures, analytics and models for the optimization of the manufacturing process, with particular focus on innovative manufacturing technologies such as High Pressure Die Casting (HPDC) and Additive.
Rsearch objectives and methods: The doctoral proposal therefore has the following objectives:

1) development of IIoT platforms suitable for managing the entire industrial manufacturing process; the platform must collect i) manufacturing process parameters, ii) process signatures (i.e., non-predefined dynamic process characteristics) from sensors and cameras, iii) product's monitoring data, iv) quality data, v) data analytics both on edge and on cloud.
2) development of new analytics for the detection of process signatures and their integration into the IIoT platforms.
3) development of techniques based on Machine Learning for i) the identification of the most significant process variables (KPI), ii) the identification of correlations between KPIs, iii) the prediction of the quality of the produced pieces, iv) the identification of the early stop conditions.
4) development of models that simulate the manufacturing process (both open and closed-loop).
5) development of closed-loop strategies that allow the real-time redefinition of process's parameters to mitigate/avoid the occurrence of defects once drifts have been identified.
Outline of work plan: I year
Objectives 1 and 2)
- Study and implementation of IIoT available solutions, with particular attention to flexible and easily adaptable solutions (e.g. SiteWhere)
- IIoT platform adaptation to an additive or HPDC context
- Integration within IIoT platform of data analytics

II year
Objective 3)
- Design and implementation of methods based on statistical analyses, machine learning (eg Random Forest, LightGBM) and deep learning (eg GNN, Transformers) for the i) identification of the most significant features (KPI), ii) prediction of defects, iii) prediction of early stop
Objective 4)
- Modeling (e.g., on Simulink) of HPDC/Additive manufacturing process through the integration of KPIs (both process parameters and signatures), and quality data (geometric, physical, mechanical)

III year
Objectives 4 and 5)
- Model consolidation
- Study of the impact of process variables on quality data
- Identification of optimal process parameters (output)
- Integration of the model's output within the IIoT platform
- Closed-loop system design
Expected target publications: IEEE Transactions on Industrial Informatics
IEEE Transactions on Industry Applications
IEEE Internet of Things Journal
IEEE Transactions on Instrumentation and Measurement
IEEE Intelligent Systems
Pattern Recognition
Robotics and Computer-Integrated Manufacturing
Required skills and competences:
Current funded projects of the proposer related to the proposal: - MANUELA – Additive Manufacturing Using Metal Pilot Line. H2020-NMBP-TR-IND-2018-2020
- STAMP – Sviluppo Tecnologico dell'Additive Manufacturing in Piemonte. 2016-2019
- DISLOMAN – Dynamic Integrated ShopfLoor operation Management for Industry 4.0. 2016-2019
- SERENA – VerSatilE pREdictive mainteNAnce platform enabling remote diagnostics and factory monitoring. 2017-2020
- AMABLE – AdditiveManufacturABLE. (H2020 project)
- FCA-POLITO – A virtual and learning plant model for energy efficiency assessment and interactive augmented visualization and prototyping
- MADEin4 – Metrology Advances for Digitized ECS industry 4.0
Possibly involved industries/companies:FCA, 2A, GE AvioAero

Title: Advancing Mobile UI Testing Techniques
Proposer: Marco Torchiano, Luca Ardito
Group website: http://softeng.polito.it/
Summary of the proposal: Testing of mobile application user interfaces consists in writing test cases that exercise the UI, and allow performing end-to-end testing of the whole app in an automated way. Tests for mobile applications are particularly challenging since they have to address a wide range of devices and platforms.
Despite the presence of several mature frameworks, UI tests show a limited adoption. This is certainly due to the significant effort in writing the tests cases, though the main barrier is represented by the inherent fragility of such kind of tests: a minor change in the UI can “break” the test requiring a significant effort in identifying the cause and then fixing the test.
Lowering the adoption barrier requires both working on existing test suites with the ability to detect fragile tests in existing suites and then techniques to fix such fragilities; in addition novel approaches to build tests ought be devised that minimize writing effort, and enable effective strategies to evolve test suites.
Rsearch objectives and methods: O1.Identification of test fragilities
This step is about the definition of techniques to detect the patterns that cause test fragility. A comprehensive taxonomy (O1.1) is the prerequisite. The automatic detection of fragilities (O1.2) can then be developed. A tool that can work as a plug-in of an IDE represents the main outcome.

O2.Test fragility mitigation
Building upon the results of O1, we need to identify remediation techniques for the fragility-inducing patterns (O2.1). The remediation techniques must be implemented in order to achieve automatic removal (O2.2). The implementation should find its place within a plug-in for an IDE.

O3.Definition of novel testing techniques
The identification of the fragility-inducing patterns represents the basis also for a gap analysis of existing techniques (O3.1). Novel techniques should be defined (O3.2) typically by leveraging the relative strengths of existing ones and using novel techniques, e.g. machine vision to generate more accurate tests.

O4.Definition of test evolution strategies
The key limitation of UI tests is represented by the need for evolving along with the application UI. A set of recommendations and strategies must be devised to guide the evolution of both existing and novel tests. In this context Machine Learning techniques can be used to provide precise recommendations for test cases adaptation.
Outline of work plan: The main activities conducted in the three years of the PhD studies are:
Task 1.1: Development of a fragility detection tool
Task 1.2: Empirical assessment of effectiveness of fragility detection
Task 2.1: Identification of the fragility mitigation techniques
Task 2.2: Development of a fragility removal tool
Task 2.3: Empirical assessment of fragility removal
Task 3.1: Analysis of limitation of existing tools and techniques
Task 3.2: Definition of a novel (integrated) testing approach
Task 4.1: Identification of a test evolution strategy
Task 4.2: Empirical assessment of test evolution strategy

The four work units will follow similar methodological approaches where
- an initial identification of the issues is conducted by means of observational methods (case studies, surveys, case control studies);
- a solution is proposed
- a proof of concept implementation is developed
- an empirical assessment is conducted typically using controlled experiments.
Expected target publications: The target for the PhD research includes a set of conferences in the general area of software engineering (ICSE, ESEM, EASE, ASE, ICSME) as well as in the specific area of testing (ICST, ISSTA).
More mature results will be published in software engineering journal, the main being: IEEE Transactions on Software Engineering, ACM TOSEM, Empirical Software Engineering, Journal of Systems and Software, Information and Software Technologies.
Required skills and competences:
Current funded projects of the proposer related to the proposal:
Possibly involved industries/companies:Several companies expressed concerns on mobile UI tests, among them three showed more interest in the specific goals of this PhD proposal: Reale Mutua Assicurazioni, Intesa SanPaolo, Engineering

Title: Testing and Reliability of Automotive oriented System-on-Chip
Proposer: Paolo Bernardi, Riccardo Cantoro
Group website: http://cad.polito.it
Summary of the proposal: Nowadays, the number of integrated circuits included in critical environments such as the automotive field is continuously growing. For this reason, semiconductor manufacturer has to guarantee the reliability of the released components for the entire life-cycle that can be up to 10 -15 years. v
The activities planned for this proposal includes efforts towards
- Study and development of innovative Fault Simulation strategies that could manage current designs complexity
- The optimization of crucial Reliability Measurements such as Burn In, System Level Test and Accelerated Operating Life Tests, in addition to circuit Repair strategies
- The design of test strategies aimed at supporting in-field error detection which are demanded by recent standards such as the ISO 26262
- The application of Machine Learning methodologies to elaborate diagnostic manufacturing volume result
- The analysis of the Reliability issues of Accelerator Cores implementing Artificial Intelligence on-chip.

The project will enable a phd student to work on a very timely subject and supported by companies currently collaborating with the research group. Expectation is to actively contribute to next generation chip reliable design.

The potential outcome of the project is both related to industrial advances and high quality publication.
Rsearch objectives and methods: The phd student will pursue the following objectives in the research activities. All of them are part of the broader scenario of the Automotive Reliability and Testing field; even if this scenario is varying in time, the current objective are looking to the most significant challenges.
Reliability measurement
- Setup a low-cost test scenario based on low-cost equipment and computational infrastructure.
1. Design and implementation of a tester able to effectively drive the test and diagnosis procedure for a large IC devices.
2. Logic diagnosis based on the collected results on a set of failing devices
3. Provide information to Failure Analysis labs for a faster identification of the root cause of a malfunctioning
- Improve the effectiveness of TEST and STRESS pattern generation through thermal/power analysis and modeling
1. Better coverage of defects such as delay and exacerbation of intermittent faults
2. Reduction of the Burn-In time
3. Better effectiveness of System Level Test procedures
- Introduction of innovative flows to increase the yield in the manufacturing process
1. Novel repair algorithms for embedded cores (i.e., memory cores)
2. False positive reduction along volume data analysis (i.e., due to power effects and excessive stress during the test phases)

In-field testing
- Development of methods for achieving high coverage of defects appearing along mission behavior as demanded by the ISO 26262
1. Key-on and runtime execution of Software-based Self-Test procedure for CPUs and peripheral cores
2. Diagnosis trace for faster failure analysis of returns from fields.
- Implementation of Infrastructure-IP and redundancy methodologies for supporting effective and efficient on-line self-test
- Setup of a environment for evaluating the impact of transient faults with dramatically reduced time
Machine learning methodologies applied to test
- Conception and implementation of machine learning methodologies to elaborate diagnostic manufacturing volume result
- Prediction of failures rates and location
- Reconstruction of failure bitmaps after a compression phase
- Prediction of the stress level through indirect measurement on chip
Reliability issues of Accelerator Cores
- HW Accelerators are becoming more and more important and their usage in safety critical field is raising questions about their reliability
- SW libraries development for such kind of embedded cores will be investigated
Outline of work plan: The working plan for the PhD student is recalling the objectives drawn in the previous sections. The order is not fixed and may vary according to the advancement during the PhD program.

1st year
1. SoC sensitivity evaluation, mission mode FIT measurement and estimation through fast methods
2. Low-Cost Tester for Delay and Transient faults, based on multi-processor cores
3. Repair strategies overview for SoC components, (especially for embedded memories)
4. Volume data analysis and implementation of ML to memory testing
5. Analysis of techniques used for guaranteeing AI hardware modules reliability

2nd year
6. Diagnosis strategies development for quickly feedback failures appearing along Burn In or causing field returns
7. Reconstruction techniques for uncomplete failure bitmaps
8. Logic diagnosis of failing devices, with the intention of providing a fault location and fault model hypothesis to failure analysts
9. Design of software libraries for AI modules

3rd year
10. Advanced techniques design and implementation for reliable and safe Automotive ICs.
11. Design of a SW platform that elaborated the failure information wrote in the DB
Expected target publications: IEEE Transactions on Computers, IEEE Transactions on Emerging Topics in Computing, IEEE Design and Test, IEEE Access
Required skills and competences:
Current funded projects of the proposer related to the proposal: In this topic, the proposer is involved in research programs funded by STMicroelectronics, Infineon, Xilinx
Possibly involved industries/companies:STMicroelectronics, Infineon, Xilinx

Title: Data-powered Truck to the Future
Proposer: Tania Cerquitelli
Group website: http://dbdmg.polito.it
Summary of the proposal: In a context of deep transformation of the entire automotive industry, starting from pervasive and native connectivity, Commercial Vehicles (heavy, light, and buses) are generating and transmitting much more data than passenger cars, with a much higher expected value, motivated by the higher costs of the vehicles and the logistic, freight, and transportation business behind them.
Many data-driven aspects are covered by such digital transformation, from data engineering and components predictive maintenance, new propulsions and batteries lifecycle management, accounting and residual value throughout the vehicle life, after-sales back-office, warranty spending optimization, and fleet and service usage optimization.
In this context, an interesting research direction is to design innovative and scalable data-driven algorithms able to translate data into knowledge to effectively support the decision-making process to turn such data-powered knowledge into real value for the market and the society.
Rsearch objectives and methods: The main objective of the research activity is transforming data coming from Connected Trucks into valuable insights to effectively create value for a variety of business use cases.
The ultimate goal of this research is to study, design and develop novel data-driven strategies to innovate and generalize the complete Knowledge Discovery from Data (KDD) generated by Automotive Commercial Vehicles.
All the analytics steps, from data collection and enrichment to knowledge extraction, validation and visualization will be addressed in an innovative fashion. Proposed solutions will be original with respect to the state-of-the-art approaches since they will address open research issues such as self-tuning, scalable and self-learning strategies. Furthermore, the application context of the Automotive Commercial Vehicles opens an additional research issue related to the design and the development of innovative and hybrid data analytics architectures to support the effectiveness and efficiency of the proposed data-driven solutions.

Proposed algorithms will be validated in different use cases with the final goals of (i) generalizing the proposed approaches and transform them into plug-and-play applications, easily exploitable in novel but similar application contexts, (ii) identifying positive business cases with a shorter return on investment to be proposed to different automotive companies, and (iii) envisioning the industrialization of the proposed data-driven methodologies.

The PhD student will also study how to tailor the proposed data-driven methodologies to other industrial vehicles such as construction equipment and agricultural machinery, and even industrial machinery adopted in the manufacturing and supply chain (Industrial IoT).
Outline of work plan: During the first year, the PhD student will study, analyze, and design automated techniques for data cleaning, machine learning modeling (predictive, clustering, reliability, optimization, image recognition), and the deployment of data-driven models in Big Data and cloud environments.
During the second year, the field of research will be widened to complex systems for monitoring and detecting model performance degradation, advanced neural networks, edge computing versus cloud computing, and different propulsion specificities (ICE, LNG, Electric, Fuel Cells).
During the third year, the PhD student will expand the underlying research and data science issues, such as data interpretation and outcome presentation and visualization, not only in the form of academic and scientific publications, but also through events and business community attendance, executives’ presentations, and deep-dive knowledge-transfer sessions.

During the 2nd-3rd years, the candidate will assess the proposed solutions in diverse applications in the context of commercial and industrial vehicles.

During all three years, the candidate will have the opportunity to cooperate in the design and development of complex data-driven solutions and proper architectures relevant for the research project and to participate in conferences presenting results of her/his research activities.

The funding company is a true believer of the technology ecosystem, and recently launched AMBG, Accenture-Microsoft Business Group: the PhD student will investigate how to leverage the Microsoft ecosystem in a proper way to support activities and prepare effective implementation for real-world applications, including Microsoft Data Suite (Data Factory, HD Insight, Data Lake, etc.) proposing comparative evaluation towards alternatives data suites.
Expected target publications: Any of the following journals
IEEE TKDE (Trans. on Knowledge and Data Engineering)
ACM TKDD (Trans. on Knowledge Discovery in Data)
ACM TOIS (Trans. on Information Systems)
ACM TOIT (Trans. on Internet Technology)
Information sciences (Elsevier)
Expert systems with Applications (Elsevier)
Engineering Applications of Artificial Intelligence (Elsevier)
Journal of Big Data (Springer)

IEEE/ACM International Conferences
Required skills and competences:
Current funded projects of the proposer related to the proposal:
Possibly involved industries/companies:Accenture S.p.A

Title: Convex relaxation based methods for gray-box model identification
Proposer: Diego Regruto (DAUIN) and Guillaume Mercère (University of Poitiers)
Group website:
Summary of the proposal: Generically speaking, data driven modeling or system identification consists in accurately estimate the parameters of dynamical models (used, e.g., to mimic the behavior of complex systems) from experimentally collected data and some prior information. Model structures encountered in system identification are often divided into three classes: white (model structure entirely based on physical equations), black (mathematical model structure with no relation with the physical equations) and gray-box models. In this project, a specific attention is paid to gray-box models, i.e., models whose structure is partially governed by some prior physical knowledge and/or first principles. Accurately estimating the parameters of a gray-box linear time-invariant state-space representation is a challenging problem especially if the number of unknowns exceeds ten, due to the fact that the model equations typically depends nonlinearly on physical parameters to be estimated. Standard nonlinear (local) optimization-based procedures often fail because the initial guesses are not in the domain of attraction of the user-defined cost function global minimum. The main goal of this project is to overcome such issues by combining linear algebra and convex-relaxation based set membership solutions which have proved their efficiency in several applied data driven modeling problems in the recent years. The activity will be developed in the context of a collaboration between the System Identification and Control (SIC) group of the DAUIN, with the Laboratory of Computer Science and Automatic Control for Systems (LCSAC) of Poitiers University lead by Prof. Guillaume Mercère.
Rsearch objectives and methods: The research activities of the System Identification and Control (SIC) group in Politecnico di Torino has been mainly focused during the last decade on the development of convex-relaxation based algorithms for solving classes of polynomial optimization problems specifically arising in the framework of set-membership estimation. In parallel, data driven modeling solutions dedicated to gray box model identification have been recently developed in the Laboratory of Computer Science and Automatic Control for Systems (LCSAC) of Poitiers University. The proposed research project is focused on combining the theoretical results and numerical algorithms developed by both research groups for proposing novel effective approaches for gray-box identification able to overcome limitation of the methods already available in the literature.

More specifically, available algorithms dealing with gray box linear time invariant state space representations are suffering from important drawbacks that can be summarized as follows:

- involve multi-stage solutions,
- require reliable black box models as starting point
- only suitable for specific model structures
- uncertainty of the estimated model cannot be easily quantified

In a nutshell, the current research project will focus on the adaptation of convex relaxation based methods to deal with gray box models (instead of black box representations mainly used until now). The main topics of the research project are summarized in the following three parts.

Part I: set-membership approach for gray box state space model identification

The main difficulty in the identification of gray-box state-space models is the nonlinear dependency on the physical parameters entering the equations. Convex relaxation techniques recently developed by the SIC group for the problem of linear and nonlinear black-box model identification will be modified and extended in order to deal with the specific nonlinear character of the gray-box estimation problem.

Part II: combining linear algebra and convex relaxation based algorithms for gray box state space model identification

Recently, a new gray box model identification technique using linear algebra and closed form solutions has been suggested by LCSAC researchers. Such a technique requires specific rank constraints to be satisfied in order to guarantee that a linear algebra based solution exists and can be determined easily. When such rank constraints are not fulfilled, the linear algebra based solution boils down to a optimization problem involving mostly one or two unknown parameters only and nonlinear optimization algorithms need to exploited to solve the problem. However, standard optimization tools typically trap in local minima leading to unacceptable solutions. In order to overcome such a problem, an alternative formulation of the optimization problem will be proposed which, being based on suitable convex relaxations, it is expected to provide accurate approximations of the global optimal solution.

Part III: computation of model uncertainty for estimated gray-box state-space models

The main goal of this part is to propose a two-stage approach for quantifying the uncertainty of a gray-box model identified by using the techniques developed in Part I and II. In the first step parameter uncertainty intervals will be computed for a black box description of the system by means of standard set-membership techniques. Then, convex relaxation techniques will be used for converting the black-box parametric uncertainty into equivalent uncertainty intervals for the gray-box model parameters.
Outline of work plan: FIRST YEAR

January 1st – June 30th :
the first six months of the project will be devoted to the study of the literature with reference to the subject of gray box model identification, linear algebra, set-membership estimation/identification, convex relaxations of polynomial optimization problems.

Milestone 1:
Report of the results available in the literature; theoretical formulation of problem and analysis of the main difficulties/critical aspect/open problems. Results obtained from this theoretical analysis of the problem will be the subject of a first contribution to be submitted to an international conference.

July 1st – December 31st:

The second part of the first years will be devoted to the problem of extending the convex relaxation techniques previously developed by the SIC group to the case of gray-box identification.

Milestone 2:
Derivation of original convex-relaxation based algorithms for gray-box identification. Results obtained in this stage of the project are expected to be the core of a paper to be submitted to an International Journal

SECOND YEAR
January 1st – June 30th :

The first half of the second year of the project will be focused on the formulation of suitable convex relaxations for the nonlinear optimization problems arising from the linear algebra approach to gray-box identification and the derivation of suitable numerical algorithms.

July 1st – December 31st:
the objective of this part will be to derive a suitable convex-relaxation based formulation of the problem of computing the uncertainty of the parameter of a gray-box model starting from the parameter uncertainty intervals of an input-output equivalent black-box model.

Milestone 3:
Critical comparison of the different algorithms derived in the first two years of research. Results obtained in this stage of the project are expected to be the core of both a conference contribution and a paper to be submitted to an International Journal.

THIRD YEAR
January 1st – December 31st:
the last year of the project will be devoted to apply the derived gray-box identification algorithms to real-world problems in the automotive field in the context of a collaboration with FCA/CRF group.

Milestone 4:
Application of the developed methods and algorithms to experimental data taken from real-world automotive problems.
Expected target publications: JOURNALS:
IEEE Transactions on Automatic Control, Automatica, IEEE Transactions on Control System Technologies, System and Control Letters, International Journal of Robust and Nonlinear Control.

CONFERENCES:
IEEE Conference on Decision and Control, American Control Conference, IFAC Symposium on System Identification, IFAC World Congress
Required skills and competences:
Current funded projects of the proposer related to the proposal:
Possibly involved industries/companies:Fiat Chrysler Automobile (FAC) and Fiat Research Center (CRF), which have collaborated with the System Identification and Control (SIC) group continuously in the last decade, will be involved in the application of the derived algorithms to different modeling, identification and control problems arising in the automotive fields.

Title: Stimulating the Senses in Virtual Environments
Proposer: Fabrizio Lamberti
Group website: http://grains.polito.it/
Summary of the proposal: The goal of virtual reality (VR) and augmented-reality (AR) is to immerse users in computer-generated, virtual environments (VEs) which could be either fully or partially digital.
Interacting with VEs poses challenges which are very unique to these technologies. In fact, an ideal VE shall let the users feel as they are physically performing a task. However, sense of immersion and presence in current VEs are still far from ideal.

Notwithstanding, there are evidences of the fact that, in order for, e.g., a VR/AR-based simulation or a training to be effective, quality of graphics, of sound, of haptic feedback, etc., do not necessarily have to be realistic, just credible. It might not be essential to recreate every possible stimulus provided from the real world, every component of the task, every sensation.

Depending on the task, level of “fidelity” required could vary significantly. A number of experiments have been carried out in various domains to evaluate actual requirements, but their heterogeneity makes the applicability of available findings to further scenarios extremely hard to achieve.

This proposal is to dig into this research field, and define a methodology for identifying the level of realism required for the experience at hand.
Rsearch objectives and methods: Over the past years, a number of techniques have been proposed to boost the level of realism of digital contents populating both VR and AR-powered immersive VEs. The aim has been to improve the way users perceive the VE by simulating an increasing number of senses, by both enhancing fidelity in reproduction of sight and hearing, as well as developing new approaches to simulate touch, smell and taste [1, 2].

In order to achieve a complete sense of immersion and presence, every stimulus provided from the real world, every sensation shall be possibly recreated – in principle – by making interaction as much faithful as possible.

Thus, when performing, e.g., a maintenance task in AR, parts and tools should have mass, feel real, and respond to the user’s action appropriately. Similarly, when simulating, e.g., an emergency situation in VR, like a tunnel or forest fire, users’ visibility and breath capability shall be reduced to make them perceive “to be there”, to be part of what is happening. However, as a matter of fact, technology is still not able to stimulate all the users’ senses in a suitable way.

To deal with the above limitations, researchers started to study to what extent the quality of contents and of the interactions users can have with them can influence the effectiveness of the task to be performed [3].

Although, in some cases, findings obtained by experimenting with a given technology and/or with a particular task may be generalized (for instance, the importance of reconstructing users’ users or providing haptic feedback has been demonstrated already under various conditions [4]), in most of the cases results are specific to a given application domain (health, industry, etc.) or task (such as, e.g., surgery preparation, rehabilitation, training, for health) [5].

The goal of this research will be to explore the large design space of simulation fidelity from many possible perspectives. The limit of various technologies will be investigated, and approaches to cope with them proposed, by taking into account both professional as well as consumer settings. Furthermore, a wide set of tasks encompassing a representative set of application scenarios will be considered, with the aim to identify best practices and guidelines applicable across domains.

To this aim, applications with a training purpose will be specifically considered through user studies since, besides qualitative indicators, quantitative metrics could be obtained as well by focusing on learners’ performance.

[1] J. Lee et al. “TORC: A Virtual Reality Controller for In-Hand High-Dexterity Finger Interaction,” ACM CHI, 2019.
[2] K. Karunanayaka et al. “New Thermal Taste Actuation Technology for Future Multisensory Virtual Reality and Internet,” IEEE TVCG, 24:4, 2018.
[3] N. C. Nilsson et al. “Waiting for the Ultimate Display: Can Decreased Fidelity Positively Influence Perceived Realism?” IEEE WEVR, 2017.
[4] F. Argelaguet et al. “The Role of Interaction in Virtual Embodiment: Effects of the Virtual Hand Representation,” IEEE VR, 2016.
[5] Z. Khan et al. “A Low-fidelity Serious Game for Medical-based Cultural Competence Eeducation, Health Informatics Journal, 25:3, 2017.
Outline of work plan: During the first year, the PhD student will review the state of the art in terms of techniques/approaches developed/proposed to deal with the issue of (level of) fidelity in VR- and AR-based environments. He or she will start with investigating new methods to deal with content- and interaction-related quality issues, experimenting with technologies that are continuously appearing on the market. He or she will then apply devised solutions to specific use cases provided by funded research projects managed by the GRAphics and Intelligent Systems (GRAINS) group, in collaboration with companies and institutions operating at the regional/national and international level. Domains of interest could encompass, e.g., energy, healthcare, robotics and autonomous systems, as well as emergency management. Results of these activities will be summarized in one or more publications that will be submitted to conferences in the field. The student will complete his/her background in AR, VR and human-machine interaction (HMI) by attending relevant courses.

During the second and third year, the work of the PhD student will build onto a few simulation scenarios related to one or more application domains (possibly selected among those mentioned above and/or related to challenges proposed by involved companies) with the aim to identify specific research questions/stress limitations concerning fidelity and, hence, design suitable solutions advancing the state of the art in the field. Activities will encompass the evaluation of users’ reaction to varying level of fidelity under different conditions, with the aim to build a set of measurable indicators and draft a collection of principle/guidelines to be considered in the design of next-generation immersive VEs. Results obtained will be reported into other conference works plus, at least, a high-quality journal publication.
Expected target publications: The target publications will cover the fields of computer graphics and human-machine interaction. International journals could include, e.g.:

IEEE Transactions on Visualization and Computer Graphics
IEEE Transactions on Human-Machine Systems
IEEE Transactions on Emerging Topics in Computing
IEEE Transactions on Learning Technologies
IEEE Computer Graphics and Applications
ACM Transactions on Graphics
ACM Transactions on Computer-Human Interaction

Relevant international conferences could include ACM CHI, IEEE VR, ACM SIGGRAPH Eurographics, etc.
Required skills and competences:
Current funded projects of the proposer related to the proposal: Topics addressed in the proposal are strongly related to those tackled in the following projects managed by the proposer:

- E2DRIVER (EU H2020), on the use of VR for training in industrial settings.
- PITEM RISK FOR (EU ALCOTRA), on the use of VR for training in emergency situations in trans-national scenarios.
- VRRobotLine (research grant in the context of a Regional project), on the use of VR for robotic applications.
- Research grant from SIPAL SpA, on the usage of AR in the engineering domain.

Activities will be carried out in the context of the “VR@POLITO” initiative and its laboratory (hub) at the Department.
Possibly involved industries/companies:- KUKA Robotics
- FCA
- Aeronautica Militare, 3° stormo Villafranca di Verona
- Regione Piemonte (Protezione Civile, AIB)

Title: Cross-Domain 3D Visual Learning
Proposer: Tatiana Tommasi
Group website: http://www.tatianatommasi.com
Summary of the proposal: Depth cameras and LiDAR sensors are important tools for every autonomous car and surveillance system for which sophisticated 3D data managing and learning algorithms are quickly becoming essential. Deep learning has achieved a great leap for 2D computer vision, and for 3D is currently demonstrating a great potential. However, the task of fully understanding the 3D real world still remains far-fetched.
How to extract information from un-structured and un-ordered point clouds and how to avoid their extremely expensive manual annotation are just two examples of the difficult and new challenges that need attention. This thesis will investigate how the 2D learning experience can be extended to 3D scenarios with a particular focus on the introduction of cross-domain adaptive methods for 3D vision applications.
Rsearch objectives and methods: A large portion of the current 3D learning literature still deals with synthetic CAD object datasets. In this setting, sample annotation is feasible and shape recognition, as well as part segmentation, show promising results. When passing from lab-controlled conditions to real-world scenarios, current 3D algorithms fail with a clear drop in performance. We propose to elaborate transfer learning, domain adaptation and domain generalization methods that allow to close the gap between synthetic and real 3D data distributions.
We will build over self-supervised learning, supported by recent evidences of its effectiveness for 3D problems [1,2,3,4]. To start with, we plan to investigate these existing approaches and run an extensive benchmark analysis. We will also combine them separately with supervised learning [5] and then study how they can be integrated together into a multi-task approach able to learn jointly on global and local 3D geometric information.
Furthermore, we plan to elaborate novel 3D generative models that can be both used for data augmentation and style transfer. These methods would also be able to leverage on multi-modal knowledge (e.g 2D vision, semantic attributes) and multi-source integration.
Finally we will move towards open set and incremental learning models able to manage conditions where target data contain more and possibly increasing number of classes with respect to a static source.

[1] Self-Supervised Deep Learning on Point Clouds by Reconstructing Space. Sauder, Sievers, NeurIPS 2019
[2] MortonNet: Self-Supervised Learning of Local Features in 3D Point Clouds, Thabet, Alwassel, Ghanem, arXiv 2019
[3] Multi-Angle Point Cloud-VAE: Unsupervised Feature Learning for 3D Point Clouds From Multiple Angles by Joint Self-Reconstruction and Half-to-Half Prediction, Han et al, ICCV 2019
[4] Unsupervised Multi-Task Feature Learning on Point Clouds, Hassani, Haley, ICCV 2019
[5] Domain Generalization by Solving Jigsaw Puzzles, Carlucci et al, CVPR 2019
Outline of work plan: M1-M6: Definition of the experimental testbed: sample collection across multiple synthetic and real 3D datasets. Literature review and implementation of existing baselines. Metric choices and extensive benchmark evaluations.

M6-M12: Implementation of a multi-task model that combines supervised and self-supervised 3D learning, extending [5]. Testing on defined benchmarks considering transfer learning, domain adaptation, domain generalization, few-shot and zero shot settings. Writing of scientific report on findings of Y1.

M13-M24: Extensions of the deep architecture obtained in Y1 to include 3D generative models. Assessment of work on the established benchmarks. Writing of scientific report on findings of Y2.

M25- M36. Extensions of the deep architecture obtained in Y2 to include open set and incremental learning conditions for 3D scenarios. Assessment of work on the established benchmarks. Writing of scientific report on overall findings of the project. Writing of PhD thesis.
Expected target publications: It is expected that the scientific results of this thesis will be reported in the top conference in the field of computer vision (IEEE CVPR, IEEE ICCV, ECCV) and robotics (ICRA, IROS, RSS). At least one journal publication is expected on one of the following international journals: IEEE PAMI, IJCV, CVIU, RAL.
Required skills and competences:
Current funded projects of the proposer related to the proposal: CHIST-ERA project BURG: Benchmarks for UndeRstanding Grasping (2019-2021)
Possibly involved industries/companies:

Title: Deep Learning of Object Parts and their Affordances for Task-Oriented Robot Grasping
Proposer: Tatiana Tommasi
Group website: http://www.tatianatommasi.com
Summary of the proposal: Can we enable artificial intelligent agents to interact with object as we do? Grasping is one of the first functional interaction that we learn and it drives our perception of the world. However, it is still a hard task for robots. Grasping rigid objects has been studied under a wide variety of conditions: the common measure of success is a check on the robot holding an object for a few seconds [1]. This is not enough. To obtain a deeper understanding of object manipulation, it is necessary to focus at the same time on part-based modeling of objects and task-oriented modelling of grasping. Moreover, to guarantee a fair benchmarking, it is crucial to define widely sharable experimental settings and reliable metrics.
The objective of the project is to extend standard robotic grasping from pick and place towards affordance reasoning and task-specific interaction [2]. This research direction requests several efforts both in the collection of part and affordance annotated 2D and 3D (depth, point cloud) datasets as well as in the development of deep learning algorithms able to do whole object and shape recognition as well as part and functional object areas segmentation. After in-lab algorithm development and simulations, practical tests will directly involve real robotic platforms.
Rsearch objectives and methods: The candidate will work mainly on machine learning techniques at the intersection between robotics and computer vision. He/She is expected to develop an in-depth knowledge of the current state of the art of robot grasping and manipulation approaches as well as a comprehension of the challenges that involve rigid and deformable objects characterization in 2D/3D computer vision and interaction with different kinds of robotic embodiments (grippers, viewpoints and RGB, RGBD acquisition modalities).
The focus of our work will be on robust deep learning methods for 2D and 3D object part detection and task-specific recognition of the related affordances, which means semantic segmentation of objects on the basis of physical attributes and potential actions performed on them.
Recognizing object parts is essential to know how and where a gripper can grasp given the constraints imposed by the task. Moreover, parts facilitate knowledge transfer to novel objects, across different sources (synthetic/real data) and grippers (e.g. 2/3 fingers, soft hands, suction cups), providing for a versatile and scalable system. The part-based approach naturally extends to deformable objects for which the recognition of relevant semantic components, regardless of the object actual deformation, is essential to get a tractable manipulation problem.
For the core of the project we plan to leverage on recently developed self-supervised learning techniques that allows to minimize the annotation efforts and have already demonstrated to provide reliable models for whole object recognition[3,4,5]. We will further challenge these techniques and develop new solutions tailored for object part recognition across modalities (rgb, depth, synthetic images) and affordance transfer across objects. Furthermore we plan to exploit human grasp first and third person videos [6] and check how the learned model transfer as initialization for robot grasping.

[1] Mousavian, Eppner, Fox, 6-DOF GraspNet: Variational Grasp Generation for Object Manipulation, ICCV 2019
[2] Do, Nguyen, Reid, AffordanceNet: An End-to-End Deep Learning Approach for Object Affordance Detection, ICRA 2018
[3] Carlucci, D'Innocente, Bucci, Caputo, Tommasi, Domain Generalization by Solving Jigsaw Puzzles, CVPR 2019
[4] Bucci, D'Innocente, Tommasi, Tackling Partial Domain Adaptation with Self-Supervision, ICIAP 2019
[5] Sauder, Sievers, Self-Supervised Deep Learning on Point Clouds by Reconstructing Space, NeurIPS 2019
[6] Nagarajan, Feichtenhofer, Grauman, Grounded Human-Object Interaction Hotspot From Video, ICCV 2019
Outline of work plan: M1-M6: A first phase of the work will be dedicate to identify the most useful existing 2D (Imagenet, Visual Genome) and 3D data testbed (Shapenet, Modelnet, ScanObject, YCB) with samples annotated with parts, physical attributes and affordance map for different kinds of final tasks. The collections will need some bridging and integration with extra annotation.

M6-M12: A second phase will focus on implementation and use of existing deep learning models for object detection, recognition, part and affordance segmentation with the assessment of their performance on the obtained cross-domain dataset. Writing of scientific report on findings of Y1.v
M13-M24: After this baseline evaluation, we plan to develop and integrate novel self-supervised deep learning strategies with symmetric multi-task as well as two-step transfer learning procedures both from images and video. Besides simple 2D self-supervision, we also plan to consider other modalities more related to robotic interaction (eg pushing for weight and friction analysis, shaking or tapping for sound). The new techniques will be analyzed on their ability of generalize across domains (synthetic to real, 2D to 3D), and effectiveness on open set conditions for the annotation of parts and grasp affordances of new object unseen at training time. Writing of scientific report on findings of Y2.

M25-M36: Finally we plan to exploit the tool and metrics developed by recent research on grasping to establish how the developed computer vision learning techniques can guide robot actions. Besides evaluation on simulated environments, the models will also be adapted to ROS coding environment and tested on real platforms for task with increasing level of difficulty: from rigid to deformable objects, from single object to cluttered scenes with presence of possible environmental constraints and in case of increasingly difficult task (object stacking, human hand over). Writing of scientific report on overall findings of the project. Writing of PhD thesis.
Expected target publications: It is expected that the scientific results of this thesis will be reported in the top conference in the field of computer vision (IEEE CVPR, IEEE ICCV, ECCV) and robotics (ICRA, IROS, RSS). At least one journal publication is expected on one of the following international journals: IEEE PAMI, IJCV, CVIU, RAL.
Required skills and competences:
Current funded projects of the proposer related to the proposal:
Possibly involved industries/companies:

Title: Cybersecurity Certifications for Hardware components and systems
Proposer: Paolo Prinetto
Group website: https://www.testgroup.polito.it
Summary of the proposal: The proposed research activity will focus on the study and development of a certification scheme for assessing the security of hardware devices.
We are living the IoT era and the world is fast moving towards what we could call the IoE (Internet of Everything) era. Everything is constantly connected to the Internet creating a huge surface for cyber attacks.
For this reason, recently the topic of cybersecurity has received increasing interest from both research (academic and not) and governments. In 2017 cybersecurity was put, by President Junker, in the second place of emergencies in Europe.
In all this it is necessary to have a system of certifications to assert the security of a component, be it hardware or software.
While a lot of work has already been done in the field of software certifications, as far as hardware certifications are concerned there is no an all-inclusive certification system, there are certifications such as Common Criteria or TEMPEST, but which focus on specific aspects and vulnerabilities.

The candidate will be asked to propose a Hardware Security Certification Schema, having identified the main automatic tests to be carried out to detect the presence or not of known vulnerabilities.
Rsearch objectives and methods: The activity carried out during the PhD period will have the aim of making it possible for the candidate, by the end of the three years, to propose a certification scheme for hardware devices in the field of computer security. In particular the candidate is asked to propose a series of tests to be performed in order to assess the security of a give device.
In order to do so, the candidate will start by analysing the existing certification schemes to understand the modalities and limitations of them, for example, Common Criteria [1] and TEMPEST [2].
As in software, also in hardware there are different types of vulnerabilities, each unique in its kind. These vulnerabilities can be exploited in a very specific way or in different manners. It is, therefore, crucial for the candidate to become familiar with and study the main hardware vulnerabilities.
The purpose of the candidate's work is to propose automatic tests for the detection of hardware vulnerabilities, given the multiplicity of these, the tests to be performed should:
1. detect vulnerabilities arising from flaws or design bugs, such as Meltdown and Spectre [3];
2. detect intentionally inserted vulnerabilities, such as Hardware Trojan [4];
3. detect vulnerabilities present as a consequence of a technological factor, such as Row Hammer [5].
The study and research activity of the candidate, although it will mainly focus on the aspects of testing, so how to discover vulnerabilities etc., will also cover more bureaucratic aspects, with which the candidate must become familiar.
With the Cybersecurity Act, the European Union has launched a project for a Cybersecurity Certification Framework with the aim of having a unified certification system at European level. This is both to encourage collaboration between member states for cybernetic defense of the Union, and to simplify the process necessary for certification (certifying a product in one member state, this will have value in all the others).
The candidate will be asked to take part in European projects, such as SPARTA [6], in which to collaborate with other members who carry out activities on similar issues.

[1] https://www.commoncriteriaportal.org/
[2] https://apps.nsa.gov/iaarchive/programs/iad-initiatives/tempest.cfm
[3] https://meltdownattack.com
[4] Tehranipoor, M., and Koushanfar, F. (2010). A survey of hardware trojan taxonomy and detection. IEEE design and test of computers, 27(1), 10-25.
[5] Gruss, D., Lipp, M., Schwarz, M., Genkin, D., Juffinger, J., O'Connell, S., ... and Yarom, Y. (2018, May). Another flip in the wall of rowhammer defenses. In 2018 IEEE Symposium on Security and Privacy (SP) (pp. 245-261). IEEE.
[6] https://www.sparta.eu
Outline of work plan: The candidate will be tutored for the whole three years in order to achieve the objective reported before.
In particular it is expected from the candidate that in the first year the following step will be performed:
1. Analysis and study of the main certification schemes adopted by the various countries and governments.
2. Analysis of the main hardware vulnerabilities and how they are exploited.
In the second year the candidate will:
1. Start the creation of a Vulnerability Database, similar to CVE [1]. This database will focus only on hardware vulnerabilities, they will be enumerated and assigned with a score, similar to CWSS [2], based on different factors, such as the ease of exploitation, the severity of the consequences and so on.
2. Start proposing automated test for assessing the security of hardware devices and the eventual presence of vulnerabilities.
The activity of the candidate will mainly focus on:
1. Finding appropriate test pattern for the various vulnerabilities
2. Definition of different level of security required based on the criticality of the application in which the device will be used
3. Definition of the type of test and control to be performed based on the security level abovementioned. Different level, different cost for the certification procedure, less stringent tests and controls.
In general it is expected by the candidate that by the end of the three year of the PhD program, she/he will become familiar with the concept of Cybersecurity Certification, what is required, which are the criticalities and so on, and to have obtained all the necessary tools to be able to be successful in its activities.

[1] https://cve.mitre.org
[2] https://cwe.mitre.org/cwss/cwss_v1.0.1.html
Expected target publications: Conferences:
- IEEE European Symposium on Security and Privacy (EurtoSP)
- Design, Automation and Test in Europe (DATE)
- VLSI Test Symposium (VTS)
- European Test Symposium (ETS)
- HOST
- Specialized Workshops
- ITASEC

Journals:
- IEEE Transactions on Dependable and Secure Computing
- IEEE Transactions on Computers
- IEEE Internet of Things
- ACM Transactions on Information and System Security
- ACM Transactions on Privacy and Security
- Design and Test of Computers
Required skills and competences:
Current funded projects of the proposer related to the proposal:
Possibly involved industries/companies:1. CINI Cybersecurity National Lab
2. Fondazione Links
3. CVCN

Title: Data mining for Digital Information Literacy
Proposer: Elena Baralis, Silvia Giordano (SUPSI Lugano, Switzerland)
Group website: http://dbdmg.polito.it/
Summary of the proposal: The research is about how late teenagers locate, access and assess online information on different topics and in different scenarios. Information search is a core life activity, be it for finding facts, facilitate decision-making or for pleasure and enjoyment. Today’s media-rich society is a highly complex information environment, and Digital Information Literacy (DIL) is central for full and aware citizenship. Innovating on methods, this research brings together digital media literacy studies and information sciences in a task-driven study to observe and analyse actual web search practices of adolescents through the collection of web search logs. Web logs will be data-mined in order to identify web search patterns, and relate them to specific control variables (socio-economic status, digital self-efficacy, declarative DIL knowledge, etc.) and search performance self-assessment. The project provides innovation under two respects: content and methods. On the first hand, the project generates novel insights on DIL, expanding existing knowledge in the domain of digital and media education. In particular, the new insights focus on (a) the difference between declarative and active literacy skills; (b) task-domain-related differences; (c) influences of socio-economic status, digital self-efficacy, interest and self-assessed DIL knowledge.
Rsearch objectives and methods: Our relationship with online information is getting more and more vital and complex. Search engines, user profiling and AI algorithms filter and shape our communications with the world and determine our information environment. Learning to effectively search online is not just a factor of academic achievement, but a key skill for active citizenship and a healthy democracy. Therefore, the main objective of this study is to understand not only where search happens, but how it happens. We can distinguish several correlated objectives to this global one:
1. This study aims at building an empirical basis to identify effective search patterns in new search settings (e.g., on mobile while travelling). This will help focusing DIL education on actual issues, factoring in actual usage, and not only on basic skills and just “guessing” how students actually search. To these purpose, Open Educational Resources (OER) for students in lower and upper secondary education will be produced.
2. At the institutional level, the insights from this study aim to support the implementation of a key part of the Media and Informatics section of the new school curricula, which can be based on sound evidence and identifiable practices.
3. Understanding how online search happens will reveal common pitfalls and ineffective behaviours. The objective is therefore to build such evidences so that search engines and information providers (not only Google, but also the ones on libraries, commercial, news or institutional websites) can benefit from them in order to improve their services (e.g., interfaces, online help, tutorials) resulting in a better online information environment.
4. Fake news spread also because of ineffective or superficial search strategies. This study aims to find insights in these strategies, in order to help tackle this crucial issue in terms of awareness-raising and education.
Outline of work plan: PHASE I (0-6 months): Data collection. The actual data collection will last 5 weeks. On day 1, the participants will receive a start-up survey, which will collect some personal information and assess control variables using already established scales. During the following 4 weeks, participants will receive an information-related task once a week via email. Data collected will then be cleaned and filtered.

PHASE II (6-18 months): Data analysis. The data analysis will be based on four main data sources: (a) participants' self-reported DIL knowledge assessed via initial questionnaire; (b) task-related web navigation logs, which represent our Information Search Behaviors (ISB) dataset; (c) participants' self assessed rating of their performance in the different search tasks; and (d) participants’ demographics and digital self-efficacy. The data mining analysis will take into account data source (b), and will consider several variables related to navigation. During this phase, the student will develop algorithms to extract identify search activity patterns that can describe specific search strategies, and combine them with other information.

PHASE III (3rd year): algorithms improvements, both in design and development, experimental evaluation with more complex and complete data collection.
Expected target publications: Any of the following journals:
ACM TOIS (Trans. on Information Systems)
IEEE TKDE (Trans. on Knowledge and Data Engineering)
ACM TODS (Trans. on Database Systems)
ACM TKDD (Trans. on Knowledge Discovery in Data)
Journal of Big Data (Springer)
IEEE TBD (Trans. on Big Data)
ACM TOIT (Trans. On Internet Technology)
OSNAM (Online Social Networks and Media – Elsevier)
APNS (Applied Network Science – Springer)
IEEE/ACM International Conferences (e.g., IEEE ICDM, ACM KDD, IEEE BigData, IEEE Asonam, TheWeb Conference, The International Conference on Complex Networks and Their Applications).
Required skills and competences:
Current funded projects of the proposer related to the proposal: Swiss NSF project LOIS: Late teenagers’ Online Information Search
Possibly involved industries/companies:NA

Title: Synthesis of Smart and Intelligent Sensors
Proposer: Enrico Macii, Andrea Calimera
Group website: http://eda.polito.it
Summary of the proposal: Intelligent sensors are the ultimate link between the “real” world (made of people, things, machines, or any other kind of connected objects) and the “cyber” reality of ICT (Information Communication Technology) infrastructures. Their function is to provide distilled information that can be used to infer the environmental conditions and/or users’ needs. This process, referred as “sensemaking”, consists of few basic steps. The raw data sampled as physical quantities (e.g. light, pressure, humidity, acceleration, radiation, etc.) are translated into an equivalent electrical signal by means of a transducer, then discretized as a digital entity by means of Analog-to-Digital Converters (ADC), stored into a local memory close to the sensor, and finally elaborated by the processing cores with the aim of extracting valuable information. The outcome is a bunch of highly informative tags used by software apps and services to understand the context and make decision accordingly in order to drive the actuators in a proper manner. All along this data-to-information conversion process, huge packets of bits are moved from a component to another, from sensors to converters, from converters to memory, from memory to CPUs, forth and backward, ceaselessly.

Since data transfers represent a major source of energy consumption and should then be avoided (or at least limited) to improve efficiency, the challenge faced by this research proposal is develop new design solutions, methods and prototype tools for the implementation low energy, high efficiency sensing systems.
Rsearch objectives and methods: The main goal is to conceive new design solutions that give ways to implement highly efficient intelligent sensing architectures where the movement of data is minimized in favour of less energy consumption. This encompasses the availability of (i) novel implementation strategies and (ii) advanced tools for the synthesis and optimization of heterogenous designs where digital and analog domains do overlap.
Within this context, the main research activities include (but are not limited to) the design of integrated circuits mapped onto CMOS technologies suited for both analog and digital components, with particular emphasis on the development of dedicated optimization algorithms and CAD tool for circuit design and modelling, as well as the analysis of experimental data ultimately aiming at providing feedback to design and manufacturing industrial teams. The activities will span the whole design hierarchy indeed, covering a multitude of aspects related to EDA tools and in particular:

Front-end design/tools

- Automatic design and optimization of new architectures that can efficiently host algorithms for data-analysis - Design techniques for concurrent area/power/performance optimization of such architectures - Development of computer-aided design (CAD) flows and hardware/software co-design

Back-end design/tools

- Circuit-level design techniques for area and power optimization of digital-in-analog components
-s CAD tools for circuits and systems integration, with emphasis on smart strategies for the portability of specifications across different design domains along the whole implementation flow
- Physical design flow (from synthesis to layout) for the optimization of non-functional properties.

All the research activities will be conducted through a close collaboration with STMicroelectronics. This will give the student access to industrial knowledge and real-life cases for proper validation of the developed techniques.
Outline of work plan: PHASE I (months 1-6):
- Study of the state-of-the-art in intelligent sensor architectures
- Study of the state-of-the-art in low power design techniques, methods and tools
- Development of some basic sensing architectures targeted towards specific application domains and assessment of their efficiency


PHASE II (months 7-30):
- Development of automatic synthesis methodologies and prototype tools (both for front-end and back-end), possibly including architecture-aware optimization algorithms.

PHASE III (months 30-36):
- Validation of the proposed design methods tools on a set of industrial applications and architectures
- Comparison with state-of-the-art approaches
- Possible manufacturing of a prototype chip for silicon-level measurements
Expected target publications: - ACM Trans. on Design Automation of Electronic Systems
- IEEE Trans. on Computer-Aided Design of Integrated Circuits and System
- IEEE Trans. on Circuits and System
- IEEE Trans. on Computers
- IEEE Trans. on VLSI Systems
- DATE: Design Automation and Test in Europe
- DAC: Design Automation Conference
- ICCAD: International Conference on Computer-Aided Design
Required skills and competences: - Basic knowledge of machine learning techniques and in particular deep learning
- Basic knowledge of speaker recognition techniques
- Basic knowledge of deep learning frameworks (Tensorflow, PyTorch or similar)
- Good programming skills and knowledge of the Python programming language
Current funded projects of the proposer related to the proposal: None
Possibly involved industries/companies:STMicroelectronics s.r.l.

Title: Implementation of Machine Learning Algorithms on Ultra-Low-Power Hardware for In-Sensor Inference
Proposer: Enrico Macii, Daniele Jahier Pagliari
Group website: http://eda.polito.it
Summary of the proposal: Semiconductor manufacturers are increasingly producing smart sensors, i.e., single integrated circuits that combine Micro Electro-Mechanical Sensors (MEMS) with analog and digital circuitry, in order to implement signal conditioning and basic digital processing on-chip. Enhancing smart sensors with on-board machine learning algorithms would permit the generation of more informative and easy-to-use outputs, e.g. an activity class (walking, cycling, etc.) for an Inertial Measurement Unit (IMU) or the count of people in a room for a thermal presence sensor. However, smart sensors have tight ultra-low-power constraints and limited silicon area budget for logic and memory, two elements that clash with the typical computational requirements of machine learning algorithms. Indeed, state-of-the-art commercial smart sensors with machine learning functionalities use very simple models such as binary decision trees, which are limited in the type of inference that they can perform and in the corresponding accuracy.

This thesis will investigate optimization techniques at the digital hardware and software level to enable the execution of more accurate machine learning models within smart sensors. The candidate will analyze different types of applications and the corresponding MEMS, in search for the optimal machine-learning model to implement in terms of accuracy/complexity trade-off and will then realize such algorithms in hardware.
Rsearch objectives and methods: The final objective of the thesis will be the development of one or more reference digital hardware architectures for implementing on-chip machine learning for a given class of smart sensors (e.g. IMUs or multi-pixel thermal sensor arrays for presence detection). Moreover, the candidate will also develop an API to program the aforementioned architecture and to exchange data with it, in order to provide some degree of user-configurability to the hardware. For example, the same hardware IMU could be used for two types of activity recognition depending on the target application.

The candidate will co-optimize the machine learning model and the underlying hardware on which it will be implemented. The most relevant model optimization will consist in the development of a methodology to select the most appropriate model family and tune the hyper-parameters to fit the target application requirements. Several machine learning algorithms will be considered depending on the target applications, from classic and statistical solutions to deep learning approaches. Besides model selection and hyper-parameters tuning, further model optimizations may include aggressive precision reduction for input data and model parameters and redundancy removal.
At the hardware level, the candidate will develop an ultra-low-power and tightly area-constrained architecture, possibly combining programmable logic and application-specific components. The programmable part will enable some degree of user-level configurability, while the application-specific part may be necessary in order to obtain a higher energy efficiency. The technological properties of the IC technology adopted in smart sensors will also be relevant for this design phase. In fact, the integration of MEMS with analog and digital components often requires the use of a bigger technological node for the digital part, which influences the effectiveness of different types of hardware optimizations (e.g. leakage optimizations) on area occupation and energy consumption.
Finally, the candidate will investigate the trade-off between flexibility (user programmability) and efficiency, looking for the best compromise between these two contrasting objectives for the considered type of devices.
Outline of work plan: PHASE I (months 1-9):
- Study of the state-of-the-art in ultra-low power hardware design
- Study of the state-of-the-art in energy-efficient machine/deep learning algorithms, with particular focus on the types of applications that can be implemented within a smart sensor
- Analysis of the different types of smart sensors available in the market, with particular focus on machine-learning related features
- Implementation of some basic applications using the existing hardware, to be later used as a baseline for comparison in terms of accuracy and efficiency.

PHASE II (months 10-18):
- Identification of a set of target applications (and corresponding MEMS types) to validate the proposed optimizations.
- Development of model selection and optimization techniques for ultra-low power and tightly area-constrained inference
- Development of the main components of the target hardware architecture and interface.

PHASE III (months 19-36):
- Validation of the proposed architecture(s) for the selected applications by means of digital/electrical simulations
- Comparison with state-of-the-art approaches
- Possible manufacturing of a prototype chip including the proposed architecture
Expected target publications: - IEEE Transactions on CAD
- IEEE Transactions on Circuits and Systems
- IEEE Design and Test of Computers
- IEEE Sensors Journal
- ACM Transactions on Embedded Computing Systems
- IEEE Transactions of Design Automation of Electronic Systems
- IEEE Journal on Internet of Things
Required skills and competences: No particular skill is required, even if of course a previous competence in quantum mechanics and QC is very welcome. Python programming is used in all available QC platforms.
Current funded projects of the proposer related to the proposal: None
Possibly involved industries/companies:STMicroelectronics s.r.l.

Title: Cybersecurity applied to embedded control systems
Proposer: Stefano Di Carlo, Ernesto Sanchez
Group website: https://www.testgroup.polito.it
Summary of the proposal: Most of today’s embedded systems include microprocessor cores able to efficiently perform an always increasing number of complex tasks. These electronic devices usually include some microprocessor cores, different memory cores, and in some cases custom logic blocks able to support the final application requirements. Manufacturer companies of embedded systems have been asked to speed up the time to market while increasing microprocessor speed, reducing power consumption, improving system security, and reducing production costs.
One of the most important issues faced by the manufacturers of embedded systems is related to the appropriate methodologies to guarantee safe and secure devices; indeed, new security issues are required to be considered in the early stages of the processor design cycle.
During this project, the Ph.D. candidate will study and provide solutions regarding how to improve the security aspects of embedded systems.
Rsearch objectives and methods: Embedded computing systems (ECS) that are at the base of several application domains (CPS, IoT, transportation, autonomous driving, etc.) must be designed having security, privacy, data protection, fault tolerance and accountability in mind from their design phase in a measurable manner. Security must be considered at all layers of an ECS design including:
- Embedded hardware vulnerabilities
- Memory Access Protection (MMU and MPU)
- Secure Debugging (HSM and Host)
- Runtime Manipulation Detection
- Secure Feature Activation (Switch Over)

In particular, the hardware is the root of trust of a full ECS; if the hardware is jeopardized, the whole system is at stake. Thus, it is imperative to enhance control and trust of hardware across the supply chain and during its operation, and this is the global objective of this Ph.D. project.

More in details, this Ph.D. project aims at investigating a new CPU-centric architecture protected against the major hardware security threats observed across different computing domains: from low-end embedded to High Performance Computing (HPC) systems, with particular emphasis on systems employed in the automotive domain.

The main result pursued by the research activities is the development of a holistic analysis framework and a complete system-level implementation strategy to guarantee security throughout the hardware design/manufacturing supply chain and the system operational lifetime through the integration of what we call an army of security custodians: dedicated monitoring and countermeasure hardware and software blocks able to remove or minimize the impact of the Integrated Circuits (IC) supply chain vulnerabilities (during design, fabrication and testing) as well as possible physical and functional side-channel attacks (for information leaking or denial-of-service) in the field.

The successful provisions of such enhanced architecture for pre- and post-deployment security vulnerability classes has the potential to minimize the cost of the secure hardware design, and thus, to deliver significant performance and power optimizations over individually applied hardware security solutions.

The combined/unified investigation of the different vulnerabilities passes through a detailed analysis and modeling of the generic system-level “perturbations” or “disturbances” occurring to the normal system operation when it is being attacked either by malicious pre-designed hardware structures or through physical or functional attacks during system operation. This in turns requires to create a unified taxonomy of the pre- and post-deployment security threats. Based on this taxonomy, the identified hardware security vulnerabilities will be prioritized, and corresponding hardware will be developed and evaluated.

Solutions will be considered at all stages of the system design process; including simulation, architectural design, and Register Transfer Level (RTL) design.

The project aspires to work on open hardware architectures such as RISC-V but also on proprietary architecture in collaboration with major players in selected application domains (e.g., automotive domain).

The current proposal will exploit the knowhow acquired during the last 10 years by the CAD Group and Test Group of Politecnico di Torino.
Outline of work plan: The proposed research activities are structured over a three years research program.

1st year: Toolchain, Taxonomy and Characterization.

The first year of activity focuses on the development of the infrastructure required to develop, analyze, evaluate and demonstrate the idea of the security custodians through the supply chain of CPU-based ECS. This includes both: (a) the physical nodes (selected hardware cores and supporting software modules) on which all protection mechanisms will be developed and analyzed, and (b) a flexible simulation/emulation platform for porting and trade-off analysis. The two pieces (actual hardware and simulation/emulation) will complement each other throughout the project duration. The impact of the hardware custodians in terms of performance, power, energy, and reliability will be evaluated by comparing several well-established benchmarks such as the SPEC, PARSEC, EEMB SecureMark, or Apollo on the original and protected chips and the ability to counteract the considered attacks will be evaluated by means of well targeted applications identified on an attack base (e.g., cryptographic algorithms). Together with this activity, the first year of the Ph.D. will be devoted to the compilation of a complete taxonomy of hardware vulnerabilities through the whole ECS supply chain. Vulnerabilities will be prioritized based on their severity in different application domain in order to drive the activities of the second year.

2st year: Design for security.

The second year of the research program is dedicated to the development of specific hardware and software custodians to counteract the most relevant hardware vulnerabilities identified during the first year. Several classes of attacks are considered including information leaking and denial of service attacks performed by exploiting hardware vulnerabilities. Particular attention will be dedicated to those attacks impacting the memory hierarchy of the system that recently gained significant attention.
Every developed custodian will be analyzed resorting to the simulation/emulation platform developed during the first year or through real prototype hardware platforms (e.g., FPGAs). The analysis will not be limited to the provided security but also extended to other design dimensions (e.g., performance, power, reliability, etc.).
The main outcome of the research activity of the second year will be the development of a library of hardware and software custodians that can be plugged in an ECS hardware design to protect against selected attacks.

3st year: Integration and demonstration.

The last year of the Ph.D. will mainly focus on integrating all solutions developed during the first two years in order to develop a complete design for security flow based on the hardware custodians.
To demonstrate the effectiveness of the developed design flow, the full architecture will be demonstrated considering a real use case. Among the different application scenarios, we will focus on embedded control systems for the automotive domain.
Expected target publications: All activities carried out during the three years research program will be the subject of scientific publication.
In particular the publication activities will start with the submission of a complete review paper on the taxonomy of relevant hardware vulnerabilities for ECS to a high-quality journal.

In parallel, results obtained from the development of the analysis framework, implementation and evaluation of the security custodians will be subject of publications at major conferences to obtain fast feedbacks from the scientific community regarding the proposed solutions.

The main conferences where the Ph.D. student will possibly publish her/his works are:
DATE: IEEE - Design, Automation and Test in Europe Conference
DAC: IEEE Design Automation Conference
ETS: IEEE European Test Symposium
ITC: IEEE International Test Conference
VTS: IEEE VLSI Test Symposium
IOLTS: IEEE International Symposium on On-Line Testing and Robust System Design
MTV: IEEE International workshop on Microprocessor/SoC Test, Security and Verification

Eventually, different aspects of the complete framework will be subject of publications submitted to high-quality archival journals such as IEEE TCAD, IEEE TVLSI, IEEE ToC.
Required skills and competences:
Current funded projects of the proposer related to the proposal: N/A
Possibly involved industries/companies:Punch Torino

Title: Speaker verification and multi-modal identity recognition
Proposer: Sandro Cumani
Group website: http://www.dauin.polito.it/research/research_groups/srg_speech_re...
Summary of the proposal: Recent years have seen increasing interests for speaker verification systems both for forensics and surveillance tasks, and in the security field as non-intrusive alternative for user authentication. The introduction of Deep Learning approaches has significantly improved the performance of speaker verification systems. However, several challenges are still open: environmental noise, short utterance duration and domain mismatch can severely impair the accuracy of speaker recognition approaches. Furthermore, new challenges related to the classification of Deep utterance representations have arisen.

The candidate will address these challenges, investigating Deep Learning (DL) approaches for utterance representation and classification that are robust to noise and effective also for short utterances. Furthermore, the candidate will investigate domain adaptation and classification techniques.

Deep Learning has also narrowed the gap between face recognition and speaker verification methodologies. For example, many classification techniques can be directly employed in both fields. The candidate will therefore also investigate the applicability of the proposed solutions to face recognition, and address the development of multi-modal and cross-modal verification approaches.
Rsearch objectives and methods: The research project will investigate Deep Learning and classification approaches for speaker verification and multi-modal and cross-modal (voice and face) recognition. The overall goal of the project is the development of methods that allow improving the verification accuracy. Based on the results, the project will also consider suitable approximations that allow deploying the developed approaches also on devices with limited computational resources.

The main objectives of the research will be:

1) Improvement of the quality of current state-of-the-art utterance representations. The work will investigate the use of Deep Neural Networks (DNN) for utterance representation, focusing on robustness to noise and duration. The work will involve the development of suitable architectures and objective functions able to improve the quality of the utterance representations.

2) Investigation of classification approaches able to improve system performance both in terms of accuracy and computational resources. The work will investigate both end-to-end frameworks based on Neural Networks and ad-hoc classification techniques that extend current state-of-the-art approaches based on Probabilistic Discriminant Analysis and Pairwise SVM.

3) Compensation of domain mismatch at extraction and / or at classification level. The work will investigate techniques for domain adaptation and unsupervised training able to reduce the accuracy degradation due to mismatches between the data used to train both the utterance extractors and back-end classifiers and the evaluation data.

4) Investigation of the effectiveness of the proposed approaches in the context of face recognition. Given the similarities of the speaker verification and face recognition tasks, the work will analyze the applicability of the developed DNN architectures, objective functions and utterance classification methods to the latter task.

5) Development of multi-modal and cross-modal biometric verification approaches. The work will investigate methods to effectively combine voice segments and face images to improve recognition accuracy. Furthermore, the work will analyze the feasibility of cross-modal verification systems that allow comparing a voice segment with a face image.
Outline of work plan: The first year will be devoted to the study of state-of-the-art Deep Learning approaches for utterance representation and classification and for face recognition. The candidate will also start implementing state-of-the-art techniques and design novel DNN architectures to improve the speaker verification accuracy.

The following years will be devoted to the development of robust methods that are able to cope with environmental noise, short duration utterances and mismatch between train and testing conditions. The candidate will design DNN architectures and loss functions aimed at improving the generalization of the utterance representations. He will also investigate both end-to-end and ad-hoc classification approaches, extending the current state-of-the-art based on PLDA and PSVM. Based on the obtained results, the candidate will analyze the applicability of the developed solutions to the field of face recognition. Additionally, he will address the topic of multi-modal and cross-modal recognition both at the feature extraction (DNN) level and at the classification level.

Alongside the aforementioned activities, the candidate will also design approximated solutions suitable for deployment in low-resource devices.

During the three years, the candidate will have the opportunity to collaborate with researchers from different countries, industrial partners and to spend a period of time abroad in leading universities. He will also participate in international challenges and evaluations, such as NIST Speaker and Language Recognition Evaluations.
Expected target publications: Possible target conferences:
- International Conference on Acoustics, Speech, and Signal Processing (ICASSP)
- Annual Conference of the International Speech Communication Association (Interspeech)
- Odyssey – The Speaker and Language Recognition Workshop

Possible target journals:
- IEEE/ACM Transactions on Audio, Speech, and Language Processing
- IEEE Signal Processing Letters
- Journal of Machine Learning Research
- Pattern Recognition
- IEEE Transactions on Pattern Analysis and Machine Intelligence
Required skills and competences: - Basic knowledge of machine learning techniques and in particular deep learning
- Basic knowledge of speaker recognition techniques
- Basic knowledge of deep learning frameworks (Tensorflow, PyTorch or similar)
- Good programming skills and knowledge of the Python programming language
Current funded projects of the proposer related to the proposal: Nuance Communication Inc. project RIC. 325/2019 (the project is expiring and will be renewed for 2020/2021)
Possibly involved industries/companies:Potentially involved companies: Nuance Communications Inc.

Title: Quantum Computing: analysis and design of quantum algorithms for engineering applications
Proposer: Bartolomeo Montrucchio
Group website: http://grains.polito.it
Summary of the proposal: Quantum Computing (QC) is a quite new research field. Up to some years ago it was mainly related to physics departments. After introduction of IBM quantum computer three years ago, it is now becoming extremely important from the engineering point of view.
Therefore this proposal is addressed to the creation of new specific competences in this framework. Main target of QC engineers will be the analysis and development of new algorithms, as well as of new technologies for building such quantum computers.
Several quantum computers are available on the market now. The most important ones are from IBM and D-Wave, but Google, Rigetti, Bull, Microsoft and others are working on this field also.
In this proposal IBM and D-Wave will be considered at first.
The PhD candidate will be required to use a strong interdisciplinary approach, since quantum mechanics must be considered together with specific computer engineering techniques like, just as examples, artificial intelligence and computer security.
Rsearch objectives and methods: Quantum computing, mainly for the work of companies like IBM, D-Wave and Google, is going to be a computer engineering challenge, since most of the algorithms developed during the last fifty years have to be almost completely redesigned, in order to fulfil the completely different hardware architecture of a quantum computer.
In the last two years IBM (in particular with Qiskit) and D-Wave have tried to create a completely new software architecture able to give to the programmer the typical APIs already available for conventional computers.
Analysis and possibly development of new algorithms will therefore be the final research objective.
This target will require studying carefully available APIs, which are continuously changing due to the impulsive improvement of QC from the hardware point of view (in 2020 D-Wave will move from about 2000 qubits to about 5000 qubits).
Therefore, the final purpose of the work will be to understand how to apply this quite new technology in the computer engineering environment. Industrial applications will be seen with particular attention, since industries will be the first to be involved in QC revolution. In particular applications like network problems, the travelling salesman problem and other optimization issues will be of interest for companies.
Outline of work plan: The work plan is structured in the three years of the PhD program:
1- in the first year the PhD student should improve his/her knowledge of quantum computing and technology, in particular since quantum mechanics and quantum computing are not seen in the previous curriculum; he/she should also follow in the first year most of the required courses in Politecnico. At least one or two conference papers will be submitted during the first year. The conference works will be presented from the PhD student himself based on the preliminary study of algorithms working on envisioned platforms.
2- In the second year the work will be both on designing and implementing new algorithms and on preparing a first work for a journal, together with another conference. Interdisciplinary aspects will be also considered. Credits for teaching will be also finalized.
3- In the third year the work will be completed with at least a publication in a selected journal summarizing the results of the algorithms implementation on platforms and technologies that will be selected as the most promising. The participation to the preparation of proposals for funded projects will be required.
Expected target publications: The target publications will be main conferences and journals related to quantum computing. Since at the moment there are only a very few of them in the computing engineering field, the choice will be done selecting, if possible, those linked to IEEE and ACM, that already started publishing specifically on QC. It is important to note that interdisciplinary aspects will be considered as fundamental, since QC is now very useful for solving problems that can come from many research fields.
Required skills and competences: No particular skill is required, even if of course a previous competence in quantum mechanics and QC is very welcome. Python programming is used in all available QC platforms.
Current funded projects of the proposer related to the proposal: A project with TIM in 2019, which will hopefully be continued in 2020. Both these projects have been done together DET department (Prof. Mariagrazia Graziano).
Possibly involved industries/companies:TIM S.p.A.

Title: Improving Quality and Reliability of Electronic Systems via accurate Defect Analysis and Manufacturing/Test Modeling
Proposer: Matteo Sonza Reorda
Group website: http://www.cad.polito.it
Summary of the proposal: This research proposal first aims at improving the ability of pinpointing and forecasting the defects arising in the most usual semiconductor technologies currently adopted by STMicroelectronics. This target will be achieved by analyzing the most common defects, and identifying their causes.
Secondly, the research proposal aims at correlating the different defects with the adopted test solutions, understanding the ability of each of the latter to detect the different categories of the former.
Finally, the proposal aims at improving the design, manufacturing and test solutions adopted by STMicroelecronics in order to improve the quality and reliability of their products.
The proposal benefits of the support of both STMicroelectronics (which will provide data coming from its products), and CNR (Istituto per la Microelettronica e Microsistemi, or IMM) , which owns a relevant knowledge on all aspects of semiconductor manufacturing, with special emphasis on new materials.
Rsearch objectives and methods: The importance of the quality and reliability of materials and industrial processes for microelectronics is a fundamental factor for the optimization of the entire integrated circuit production cycle.
Achieving an acceptable yield requires the controlled usage of materials with low density of defects and the application of lithography and patterning methodologies able to achieve the desired structural / electronic properties while minimizing the times and costs for testing.
In this research proposal we aim at the development and experimental application of advanced modeling and characterization methodologies for optimized growth and handling of materials with particular relevance in the field of microelectronics.
As reference materials we will consider all those used in current semiconductor technologies for applications in key sectors such as IoT, Automotive and Industry.
The research activity will focus on (a) the development of advanced atomic-level algorithms for the modeling of bulk growth processes and epitaxial layers, (b) the application of attack / deposition simulations used for the realization of microelectronic devices. In both cases, state of the art numerical methodologies will be implemented to achieve space / time scales that reflect the respective industrial processes.
A key objective will be to obtain computation techniques that assist in the control of defects (which may form during manufacturing processes) and in the assessment of new technologies. In fact, the innovative characteristics of the simulation techniques to be implemented will include the possibility of
- predicting the formation, evolution and electrical behavior of defects (or materials in the presence of defects) within a physical approach;
- assessing the impact of defects on product performance from the level of the device to that of architecture and circuit;
- providing support for the development of testing methodologies that correlate electrical failures to the presence of defects.
The modeling activity will be instrumental to devise a set of optimized experiments and processes that will lead to the creation of prototypes of materials and devices taking advantage of the infrastructure and facilities of STMicroelectronics. These devices will be further characterized experimentally in the laboratories of the institute for microelectronics and microsystems (CNR-IMM) in Catania also with atomic resolution analysis techniques.
The PhD activities will also include the development of quality and reliability improvement strategies for electronic devices, examining the various types of physical defects at a microstructural level through atomic resolution microscopy techniques and dedicated computational experiments. The continuous comparison with the state of the art of international research will allow the critical evaluation of the results obtained.
Outline of work plan: The PhD student will mainly perform his research activities at the Catania site of STMicroelectronics, although he will also spend some time at Politecnico di Torino – Dipartimento di Automatica e Informatica and at CNR IMM. This will allow the student to benefit from the knowledge and expertise existing at the 3 Institutions.
The plan for the activities to be performed by the student is the following:

Year 1: Analysis of the state of the art and identification of the most relevant issues from the industrial point of view. In particular, the following goals will be pursued in this phase:
- Selection of a meaningful set of representative industrial cases of study
- Analysis of the test, diagnosis and defect modeling techniques for advanced products and technologies
- Study of the performance of the known techniques on the selected cases of study, highlighting the strong and weak points of each solution
- Identification of the main criticalities from the industrial point of view.

Year 2: Devising new techniques for improving the quality and yield of the selected products
- Identification of strategies to improve quality and reliability by acting on design, manufacturing and test
- Planning the experimental activities required to validate the devised solutions
- First dissemination of the preliminary results achieved, mainly targeting international conferences in the area.

Year 3: Development of suitable models and their assessment, in order to make the proposed solutions usable in the industrial practice
- Further experimental activities
- Analysis of the gathered results
- Extended dissemination activities, targeting not only conferences, but also journals and books.
Expected target publications: Papers at the main international conferences related to test and reliability (e.g., ETS, ATS, VTS, IOLTS, ITC).
Papers on the main IEEE journals related to design, test and reliability (e.g., IEEE Design and Test, IEEE Transactions on VLSI, IEEE Transactions on CAD, IEEE Transactions on Reliability, IEEE Transactions on Computers), IEEE Transactions on Device and Materials Reliability
Required skills and competences: Basic skills in electronic circuit design and integrated circuits manufacturing processes.
Current funded projects of the proposer related to the proposal: The research proposal benefits of the support of STMicroelectronics and CNR and aims at improving the skills of a STMicroelectronics employee under the scheme known as dottorato industriale.
Possibly involved industries/companies:STMicroelectronics and CNR

Title: Formal and non-formal learning in virtual environments
Proposer: Fabrizio Lamberti
Group website: http://grains.polito.it/
Summary of the proposal: Virtual reality and related technologies are regarded as game changers in education, since they can allow users to engage in realistic learning experiences that could either complement existing ones, or even offer ways to replace them, when not feasible.
Virtual training already proved to be very effective in various professional, non-formal education contexts. However, several barriers are still delaying a widespread diffusion in formal settings, like schools and universities.

A major limiting factor is represented by the lack of contents; authoring tools targeted to end-users as well as common design methodologies need to be developed and validated. A further shortcoming concerns readiness of technology, which still asks researchers to investigate ways to enable a seamless degree of interaction (with contents, but also with instructors and other learners) capable to let the users focus on actual learning objectives rather than struggle with technological issues. Finally, ways to boost engagement and motivation, e.g., in self-/distant-learning, scenarios involving virtual instructors need to be identified.

The goal of this research is investigating ways to deal with such challenges, building upon the best practices developed in professional settings and blending them with needs expressed by other stakeholders to further push adoption in both the scenarios.
Rsearch objectives and methods: The use of virtual reality and related technologies is now quite common in non-formal education. A number of success stories show how various industries are already training their workforce in virtual settings, medical doctors simulate surgeries before intervening on patients, emergency responders are prepared to face critical situations, etc.

When considering school- and university-level education, situation is extremely jeopardized. A few universities are investing large amounts of money to set up innovative virtual labs and create ad-hoc contents for their courses, while some visionary schools are experimenting with third-party tools, making them suit their needs. In the middle, some big players started to create their hardware and software solutions to bridge this gap.

Differences between formal and non-formal education can be ascribed, among others, to the diverse needs motivating investments, to the kinds of users involved, and to the way contents are delivered.
A company may invest in virtual education since its training needs cannot be addressed using traditional approaches, e.g., because too costly or dangerous; it may target a small set of workers, and could provide them with high-end technologies. Universities and schools may be interested in using these technologies to improve their current learning methods; they should provide the same opportunities to all the students, who may be quite numerous and equipped with different devices.

In fact, most of the learning experiences have been created to tackle specific, often school/university- or subject/course-specific needs, and did not go beyond the given context. Hence, they are frequently developed from scratch with general-purpose tools, and devising ad-hoc methodologies that try to transfer traditional education contents and paradigms from a real to a virtual environment. Although in some cases their effectiveness has been evaluated, standard methods are lacking and large-scale validity studies have still to be performed.

Despite the higher success, approaches adopted in professional scenarios do not differ significantly, and the most serious challenges are indeed shared among the above domains.

In fact, although companies could re-use, e.g., digital assets created for their operations, the lack for consolidated design methodologies and of authoring tools dedicated to learning make contents creation a very time-consuming tasks, generally leading to non-interoperable and hardly re-usable tools. The identification of the best strategy for dealing with a particular learning need is also something very crucial. Finally, learners’ interaction with the virtual environment and with other subjects is key to the success of any experience, especially in distance-learning scenarios. A poor quality of contents or of stimulations could harness the users’ sense of immersion and presence, whereas unintuitive and unnatural ways to communicate could limit the learning potential. On the contrary, the use of virtual agents capable to motivate the learners, and the adoption of objective (self-) evaluation methods could boost participation and engagement.

This research will tackle this scenario and will propose ways to address key challenges above through a holistic, user-centered approach, focusing in particular on issues concerning the design and creation of contents, and considering interaction factors capable to boost their effectiveness.
Outline of work plan: During the first year, the PhD student will review the state of the art in terms of techniques, approaches and tools developed to deal with the creation and the delivery of virtual learning experiences. He or she will study use cases being investigated within funded research projects managed by the GRAphics and Intelligent Systems (GRAINS) group also in the context of the VR@POLITO lab, in collaboration with companies and institutions operating at the regional/national and international level. Domains of interest could encompass, e.g., energy, medicine, manufacturing, and emergency management. He will start applying the achieved competences in the design and development of innovative learning experiences, touching with hands challenges charactering the considered scenarios. Results of these activities will be summarized in one or more publications that will be submitted to conferences in the field. The student will complete his/her background in virtual reality, and human-machine interaction and learning technologies by attending relevant courses.

During the second and third year, the work of the PhD student will move to addressing the specific issues pertaining virtual learning contents creation and interaction, by exploring strategies enabling a standardized and simplified transfer of traditional learning contents to virtual environments, and investigating factors capable to boost user experience in interacting with them. He or she will continue to consider use cases provided by group’s/lab’s projects, but he will blend them with education needs expressed by researchers and professors operating locally at Politecnico di Torino and at other schools/universities the group and VR@POLITO are collaborating with in this specific context (DISEG, DENERG, DISAT, Ontario Tech University, Universidad Politécnica de Madrid, University of Hildesheim, Accademia delle Scienze, etc.), with the aim to foster replicability of approaches from the non-formal to the formal education domain. Results will be reported into other conferences plus, at least, a high-quality journal.
Expected target publications: The target publications will cover the fields of computer graphics, human-machine interaction, and learning technologies. International journals could include, e.g.:
- IEEE Transactions on Visualization and Computer Graphics
- IEEE Transactions on Human-Machine Systems
- IEEE Transactions on Emerging Topics in Computing
- IEEE Transactions on Learning Technologies
- IEEE Computer Graphics and Applications
- ACM Transactions on Graphics
- ACM Transactions on Computer-Human Interaction
- IEEE Transactions on Learning Technologies

Relevant international conferences could include ACM CHI, IEEE VR, IEEE Frontiers in Education (FIE), IEEE Global Engineering Education Conference (EDUCON), etc.
Required skills and competences: Very good knowledge of tools for the creation of virtual experiences and very good programming skills, previous experience with the creation of virtual training contents.
Current funded projects of the proposer related to the proposal: Topics addressed in the proposal are strongly related to those tackled in the following projects managed by the proposer:
- E2DRIVER (EU H2020), on the use of VR for training in industrial settings.
- PITEM RISK FOR (EU ALCOTRA), on the use of VR for training in emergency situations in trans-national scenarios.
- VRRobotLine (research grant in the context of a Regional project), on the use of VR for robotic applications.
- Research grant from SIPAL SpA, on the usage of AR in the engineering domain.

Activities will be carried out in the context of the “VR@POLITO” initiative.
Possibly involved industries/companies:Companies/Entities possibly involved in the proposal:
- KUKA Robotics
- SIPAL
- Aeronautica Militare, 3° stormo Villafranca di Verona
- Regione Piemonte (Protezione Civile, AIB)
- Other Departments of Politecnico di Torino and other universities/schools

Title: Design of Secure Computer Hardware Architectures
Proposer: Paolo Prinetto
Group website: https://www.testgroup.polito.it
Summary of the proposal: In recent times, security has taken on the role of essential requirement for any type of computing machine, from the server clusters of large corporates to end-points represented by smart objects surrounding our daily lives. The communication encryption systems, which for years have been - and are still - considered the basis against attacks and intrusions, are fundamental but not sufficient. Computing systems need to implement defenses that are not limited to application data protection, but scale more and more towards the lowest levels of abstraction, until they reach the hardware on which programs and services run. If the concepts of virtualization, isolation, supervision, memory protection and secure execution are included in the design paradigm of microprocessors and hardware components, the systems would benefit from an architectural protection that goes beyond all possible vulnerabilities of software or communication protocols.
The PhD proposals aims at proposing, studying, and developing architectural solutions for computing hardware, able to guarantee predefined level of security to systems. The work will target examples of open core architecture suitable for research purposes (e.g., RISC-V) to reason about, elaborate and test such solutions.
The topic is so huge that two candidates are foreseen.
Rsearch objectives and methods: The research objectives will be:

1) Analyzing security issues by considering different vulnerabilities and possible attack surfaces. Software vulnerabilities are such that information on which the confidentiality, integrity and availability constraints apply, are instead leaked, corrupted or inhibited. In the design of a security-oriented computing hardware, these constraints must be considered as potentially not guaranteed by the upper layers, and therefore to be safeguarded already at the lowest possible abstraction level.

2) Several different solutions will be considered, depending on the type of vulnerability and the level of criticality of the system that uses the devices. These will include:
a. solutions preventing the leakage of information between different security domains, defined and enforced in hardware;
b. solution preventing the hijacking of the behavior of the system into malicious actions (e.g., Arbitrary Code Execution);
c. solutions to detect potentially suspect behavior and warn the execution environment about the possible threat;
d. solutions aimed at confining the potential malicious execution outside the access to critical resources, without affecting the overall execution environment;
e. solutions exploiting trust zones, i.e., regions of the computing hardware used to run critical code or to store critical information;
f. solutions exploiting coordination between existing features of processors to actuate security policies.

While the results obtained during the research period are expected to be general and hold for any platform, the work during the PhD thesis will explicitly focus on the RISC-V open architecture [1], a project started in 2010 by the University of California and contributed by primary actors within academies and companies.

[1] https://riscv.org/
Outline of work plan: The following steps will be taken:

1. Comprehensive analysis of:
* software vulnerabilities and attacks
* proposed hardware-based countermeasures
* their effectiveness;

2. Identifying a proper set of case studies;

3. Designing a set of solutions exploiting the alternatives summarized at point 2) of the previous section on Research objectives;

4. Developing, comparing, evaluating, gathering experimental results from the implementation for the different solutions;

5. Leveraging on the acquired experience, defining:
* a set of rules to be adopted to design secure hardware architectures;
* Design-for-Trust and Design-for-Security rules to be followed when designing new systems.

The candidates are expected to perform:
- steps 1-2 in the 1st year,
- steps 3-4 in the 2nd year,
- steps 5 in the 3rd year.

The candidates will be tutored to fully reach the above-mentioned research objectives.
Expected target publications: Conferences:
- IEEE European Symposium on Security and Privacy (EurtoSandP)
- Design, Automation and Test in Europe (DATE)
- VLSI Test Symposium (VTS)
- European Test Symposium (ETS)
- HOST
- Specialized Workshops
- ITASEC

Journals:
- IEEE Transactions on Dependable and Secure Computing
- IEEE Transactions on Computers
- IEEE Internet of Things
- ACM Transactions on Information and System Security
- ACM Transactions on Privacy and Security
- Design and Test of Computers
Required skills and competences:
Current funded projects of the proposer related to the proposal:
Possibly involved industries/companies:CINI Cybersecurity National Lab
Fondazione Links
Blu5 Lab

Title: Security Enhancements in Embedded Operating Systems
Proposer: Paolo Prinetto
Group website: https://www.testgroup.polito.it
Summary of the proposal: For some years now we are experiencing an increasing growth in the adoption of the Internet of Thing paradigm, devices always connected that collect, process and exchange information. These devices, known as embedded systems, are mainly characterized by their low power consumption and limited computing capabilities.
The need to reduce the complexity in the development of embedded applications has led to the birth of several Operating Systems (OSs), Embedded Oss, each one tailored for specific needs.
The development of these OSs has never focused primarily on security aspects, but recent cyber-attacks, such as the Mirai Botnet, have shown how even these small systems can pose a considerable risk.

The candidate, during the PhD, will focus on researching and developing solution for enhancing the security of Embedded Operating Systems. These solutions must be capable of guaranteeing pre-defined security levels, even in the presence of vulnerabilities of different nature (hardware and/or software), known or even not yet revealed. Moreover, these solutions pose several challenges with respect to solution adopted in normal OSs such as the limited computational power of the devices, the real time nature of embedded application and the need of a seamless integration in the system.
Rsearch objectives and methods: The research objectives will be:

1. Analyzing security issues in Embedded Operating Systems, by considering different type of vulnerabilities and attack vectors.

2. Several different solutions will be considered, depending on the type of vulnerability and the level of criticality of the system that uses the devices. These will include:
a. solutions based exclusively on software and aimed at preventing malicious attackers from exploiting known vulnerabilities;
b. solutions to prevent or mitigate attacks that aim at gaining control over the system (e.g., Arbitrary Code Execution [1]);
c. solutions aimed at tolerating behavior byzantine in the case of complex systems with many interacting devices [2];
d. solutions aimed at imposing and guaranteeing strong isolation methodologies of the underlying hardware resources;
e. solutions exploiting a proper subset of the built-in facilities today available in some of the most advanced processors in order to integrate them in the OS architecture;
f. solutions aimed at addressing particular needs of IoT devices, such as secure boot and remote update;
g. Solutions that exploit the integration and cooperation of different hardware security devices, such as FPGAs and Smart Cards, in order to harden the security of the OS.
While the results obtained during the research period are expected to be general and hold for any platform, the work during the PhD thesis will explicitly focus on the SEcube™ open security platform [3] that will be made available to the candidates thank to a collaboration with Blu5 Labs and with the CINI Cybersecurity National Lab.
The SEcubeTM (Secure Environment cube) platform is an open source security-oriented hardware and software platform constructed with ease of integration and service-orientation in mind. SEcube™ is a 3D SiP (System-in-Package) which embeds three components:

3. a STM32F4 microcontroller unit, equipped with an ARM Cortex-M4 processor;

4. a programmable hardware device (FPGA);

5. an EAL 5+ certified Smart Card.

[1] Carlini, Nicholas, and David Wagner. "{ROP} is Still Dangerous: Breaking Modern Defenses." In 23rd {USENIX} Security Symposium ({USENIX} Security 14), pp. 385-399. 2014.
[2] Lamport, Leslie, Robert Shostak, and Marshall Pease. "The Byzantine generals problem." In Concurrency: the Works of Leslie Lamport, pp. 203-226. 2019.
[3] https://www.secube.eu/
Outline of work plan: The following steps will be taken:

1. Comprehensive analysis of:
* software vulnerabilities and attacks
* proposed security solution adopted in modern Embedded OSs
* their effectiveness;

2. Identifying a proper set of case studies;

3. Designing a set of solutions exploiting the alternatives summarized at point 2) of the previous section on Research objectives;

4. Developing, comparing, evaluating, gathering experimental results from the implementation for the different solutions;

5. Leveraging on the acquired experience, defining a set of rules to be adopted to design secure Embedded Operating Systems;

The candidate is expected to perform:
- steps 1-2 in the 1st year,
- steps 3-4 in the 2nd year,
- steps 5 in the 3rd year.

The candidate will be tutored to fully reach the above-mentioned research objectives.
Expected target publications: Conferences:
- IEEE European Symposium on Security and Privacy (EurtoSandP)
- Design, Automation and Test in Europe (DATE)
- VLSI Test Symposium (VTS)
- European Test Symposium (ETS)
- HOST
- Specialized Workshops
- ITASEC

Journals: - IEEE Transactions on Dependable and Secure Computing
- IEEE Transactions on Computers
- IEEE Internet of Things
- ACM Transactions on Information and System Security
- ACM Transactions on Privacy and Security
- Design and Test of Computers
Required skills and competences:
Current funded projects of the proposer related to the proposal:
Possibly involved industries/companies:CINI Cybersecurity National Lab
Fondazione Links
Blu5 Lab

Title: Self-supervised cross-domain activity classification from multiple information channels
Proposer: Barbara Caputo
Group website: https://scholar.google.com/citations?user=mHbdIAwAAAAJ
Summary of the proposal: Recognizing human actions from videos is one of the most critical challenges in computer vision since its infancy. An open challenge in the field is how to build systems dynamically robust in the face of (more or less) abrupt changes in the imaging conditions with respect to what seen by the system during training. This problem, known in the literature as cross-domain visual classification, is especially challenging in the video domain, as the amount of incoming frames at test time is such to make it impossible the use of approaches requiring any form of supervision. This PhD will investigate this issue by leveraging over self-supervision, i.e. the ability to learn about domain invariance through auxiliary tasks, with the self-supervision happening across different information channels. We will consider both standard RGB videos, where the motion and appearance channels can be seen as different information streams, as well as videos acquired with multiple modalities, from audio-visual signals to RGB-D/RGB-RI sequences. While the work will initially consider specific data modalities, the ambition is to arrive by the end of the thesis to principled algorithms that can work in any setting.
Rsearch objectives and methods: With the growing popularity of portable image recording devices as well as online social platforms, internet users are generating and sharing an ever-increasing number of videos every day. According to a recent study, it would take a person over 5 million years to watch the amount of video that will be crossing global networks each month in 2021. Therefore, it is imperative to devise systems that can recognize actions and events in these videos both accurately and efficiently. To this end, an open challenge is that video analysis systems will always be trained on a data domain different from the one that will be seen at test time: imaging conditions in term of lighting, recording device and post-processing will in general differ, as well as the location where the video is going to be acquired, even though recording similar activities. This problem, known in the literature as cross-domain classification, has long been studied in the image classification domain and has recently started to attract attention also in the activity classification community.
An open issue in video processing and understanding is the sheer amount of data to be processed: while the growth in scale of the training data has enabled more effective end-to-end training of deep models, greatly advancing the field, this progress has come at a high cost in terms of time-consuming manual annotations, that become unfeasible in an 'adaptation on the fly' scenario.

This PhD will investigate how to leverage over recent advances in self-supervised object classification learning [a, b] in the context of cross-domain video analysis for activity classification. Taking inspiration from recent works on self-supervised audio-video processing [c,d], the project will investigate how to solve auxiliary tasks across various information channels from videos in a way that makes the solution of such tasks consistent across information channels and gains robustness from it. While obvious settings where to study the problem are audio-visual data and multi-modal videos (such as RGB-D or RGB-RI sequences), any video can be seen as the combination of appearance and motion, and studies in a two stream architecture dealing separately with frames and optical flow, and possibly skeletons, will be considered. The main technical objective will be the development of a principled deep architecture able to work on any information channel and to learn what self-supervised task should be used and how, in a given pre-trained architecture. The methods developed will be tested on standard benchmark databases and in public challenges, so to make it possible to assess progress with respect to the state of the art in the field.

[a] F. M. Carlucci, A. D’Innocente, S. Bucci, B. Caputo, T. Tommasi. Domain generalization by solving jigsaw puzzles. Proc. IEEE CVPR 2019.
[b] A. D’Innocente, F. Cappio Burlino, S. Bucci, B. Caputo, T. Tommasi. Learning to generalize one sample at a time with self-supervision. Arxiv 1910.03915, 2019.
[c] B. Korbar, D. Tran, L. Torresani. Cooperative learning of audio and video models for self-supervised synchronization. Proc. Nips18
[d] H. Alwassel et al. Self-supervised learning by cross modal audio-video clustering. Arxiv 1911.12667, 2019.
Outline of work plan: The research workplan will be articulated as follows:

M1-M4: Implementation and testing of [a,b,c,d] on reference benchmark databases in conjunction with off-the shelf state of the art in video activity classification.

M5-M12: Design and implementation of a cross-domain loss for adaptive activities classification with cross-modal self-supervised auxiliary tasks. Assessment of work on the on the established benchmarks. Testing of the effectiveness of various self-supervised tasks for the activity classification from multimodal/multi-stream video setting. Writing of scientific report on findings of Y1.

M13-M24: Design and implementation of a computationally efficient cross-domain loss for adaptive activities classification with cross-modal self-supervised auxiliary tasks. Assessment of work on the on the established benchmarks. Testing of the effectiveness of various self-supervised tasks for the activity classification from multimodal/multi-stream video setting. Writing of scientific report on findings of Y2.

M25-M36: Implementation of final deep architecture incorporating the best results from Y1 and Y2, with principled learning of the optimal self-supervised task and reduction of the internal parameters of the network. Assessment of work on the established benchmarks. Writing of scientific report on overall findings of the project. Writing of PhD thesis.
Expected target publications: It is expected that the scientific results of the project will be reported in the top conference in the field of computer vision (IEEE CVPR, IEEE ICCV, ECCV). At least one journal publication is expected on one of the following international journals: IEEE PAMI, IJCV, CVIU.
Required skills and competences: The successful candidate should have a degree in computer science, computer engineering, mathematics or physics, with previous experience in machine and deep learning. A master thesis work on machine/deep learning applied to visual data, and previous publications in research topics related to the project are desirable.
Current funded projects of the proposer related to the proposal: ERC RoboExNovo
Possibly involved industries/companies:Inxpect

Title: Unsupervised cross domain detection and retrieval from scarce data for monitoring of images in social media feeds
Proposer: Barbara Caputo
Group website: https://scholar.google.com/citations?user=mHbdIAwAAAAJ
Summary of the proposal: Social media feed us every day with an unprecedented amount of visual data. Conservative estimates indicate that roughly 10^1-10^2M unique images are shared everyday on Twitter, Facebook and Instagram. Images are uploaded by various actors, from corporations to political parties, institutions, entrepreneurs and private citizens. For the sake of freedom of expression, control over their content is limited, and their vast majority is uploaded without any textual description of their content. Their sheer magnitude makes it imperative to use algorithms to monitor, catalog and in general make sense of them, finding the right balance between protecting the privacy of citizens and their right of expression, while fighting illegal and hate content. This in most cases boils down to the ability to automatically associate as many tags as possible to images, which in turns means determining which objects are present in a scene. This PhD project will develop algorithms to automatically tag images from social media feeds and classify them with respect to their content, developing algorithms for detection and content-based image retrieval able to work robustly when it is not possible to make strong hypothesis on the visual domain where the incoming test image has been acquired.
Rsearch objectives and methods: Object detection and content-based image retrieval have been largely investigated since the infancy of computer vision. With the shift from shallow to deep learning, several successful algorithms have been proposed in both fields. They mostly assume that training and test data come from the same visual domain. While this is a reasonable assumption in several applications, some authors have started to investigate the more challenging yet realistic scenario where the training data come from a visual source domain, and the learned algorithm is deployed at test time in a different target domain. This setting heavily relies on concepts and results from the domain adaptation literature. In particular, most works cast the problem in the unsupervised domain adaptation framework, where the detector/retrieval system has access at training time to annotated source data and unsupervised target data, from which it learns how to adapt across the two domains.

This approach is not suitable, neither effective, for monitoring social media feeds. Consider for instance the scenario where there is an incoming stream of images from various social media and the algorithm is asked to look for instances of the class bicycle. The images come continuously, but they are produced by different users that share them on different social platforms. Hence, even though they might contain the same object, each of them has been acquired by a different person, in a different context, under different viewpoints and illuminations --in other words, each image comes from a different visual domain, different from the visual domain where the detector has been trained. This poses two key challenges to current cross-domain algorithms: (1) to adapt to the target data, these algorithms need first to collect feeds, and only after enough target data has been collected they can learn to adapt and start performing on the incoming images; (2) even if the algorithms have learned to adapt on target images collected from the feed up to time t, there is no guarantee that the images that will arrive from time t+1 will come from the same target domain.

The main objective of this PhD thesis is to investigate this largely unexplored scenario, developing algorithms able to adapt from scarce, non annotated data coming from varying target domain, in the object detection and image retrieval settings. The two scenarios will be analyzed together both for their relevance in the social feed monitoring, both for the cross-fertilization that historically have brought one on the other: indeed, object detectors are used to tag images, and the content of images primes detectors offering contextual cues. We will explore the use of self-supervised approaches for adaptation, that recently have proved to be highly effective in domain adaptation tasks [a], and will explore the challenge of moving adaptation from the feature to the metric level, a necessary step for dealing effectively with retrieval over large scale training data.

[a] F. M. Carlucci, A. D’Innocente, S. Bucci, B. Caputo, T. Tommasi. Domain generalization by solving jigsaw puzzles. Proc. IEEE CVPR 2019.
Outline of work plan: The research workplan will be articulated as follows:

M1-M4: Implementation and testing of [a] on reference benchmark databases in conjunction with off-the shelf state of the art in object detection and image retrieval; acquisition and annotation of a social medial database; implementation and testing of relevant baselines in the literature.

M5-M12: Design and implementation of a cross-domain loss for adaptive object detection from scarce data in a social media monitoring scenario. Assessment of work on the on the established benchmarks. Testing of the effectiveness of various self-supervised tasks for the detection setting. Writing of scientific report on findings of Y1.

M13-M24: Design and implementation of a cross-domain loss for adaptive image retrieval from scarce data in a social media monitoring scenario, acting at the metric level rather than at the feature level. Testing of the effectiveness of various self-supervised tasks for the retrieval setting. Assessment of work on the on the established benchmarks. Writing of scientific report on findings of Y2.

M25-M36: Implementation of final deep architecture for the monitoring of social media feeds incorporating the best results from Y1 and Y2, with principled reduction of the internal parameters of the network. Assessment of work on the established benchmarks. Writing of scientific report on overall findings of the project. Writing of PhD thesis.
Expected target publications: It is expected that the scientific results of the project will be reported in the top conference in the field of computer vision (IEEE CVPR, IEEE ICCV, ECCV). At least one journal publication is expected on one of the following international journals: IEEE PAMI, IJCV, CVIU.
Required skills and competences: The successful candidate should have a degree in computer science, computer engineering, mathematics or physics, with previous experience in machine and deep learning. A master thesis work on machine/deep learning applied to visual data, and previous publications in research topics related to the project are desirable.
Current funded projects of the proposer related to the proposal: CINI-VIDESEC (3Y, 1.2M €)
Possibly involved industries/companies:PCDM -DIS

Title: Advanced Programming Techniques for Massively-Parallel and Embedded Processors
Proposer: Stefano Quer and Giovanni Squillero
Group website: http://fmgroup.polito.it/quer/
Summary of the proposal: Mass-market computing systems combining multiprocessors boards, multi-core CPUs and many-core GPGPUs (General Purpose Graphical Processing Units) have made terascale computing available to end users. At the same time, the embedded systems pervasive in everyday life are characterized by a computational power that is reduced considering individual units but can be significant looking at the system as a whole.

In this scenario, one of the major changes in the computer software industry has been to move away from the sequential programming paradigm, creating multi-core, many-core and cooperative programming approaches. Unfortunately, all parallel programming paradigms have their own limitations. Native parallel threading approaches allow high control but are very low-level and platform dependent. High-level architecture-independent models (such as Thread Building Block, OpenMP, and Cilk) enable easy integration into existing applications but have many efficiency issues and limitations. GPGPU applications, written in CUDA or OpenCL, are well-suited to address problems that can be expressed as data-parallel computations but lack in generality. Moreover, many cooperative approaches suffer from communication overheads, such as memory transfer times, and have a very limited portability.

This proposal concentrates on using new programming paradigms to solve problem with strong theoretical bounds in three different fields, namely graph-based modeling, CAD, and learning agents. In these scenarios, we will firstly try to parallelize existing sequential algorithms without recasting them into parallel models. Then, we will analyze compression techniques, which are crucial for representing huge and massive data structures. Finally, we will study techniques that scale well. The final goal of the project will be the development of applications running faster or able to solve bigger problems on the same hardware architectures.
Rsearch objectives and methods: This project will explicitly focus on the limitations of the current state-of-the-art approaches, considering applications in the following three areas.

Efficient Parallel Graph Algorithms
Graph Data Modeling sets a new standard for visualization of data models based on the property graph approach. For that reason, there has been significant recent interest in parallel graph processing due to the need to quickly analyze very large graphs. To give an idea, Facebook has billions of nodes and trillions of edges, and the largest publicly available real-world graph today, is the Hyperlink Web graph with over 3.5 billion vertices and 128 billion edges. Unfortunately, even if graphs of this size can fit in the memory of a single commodity multicore server, many graph codes have been designed for distributed memory or external memory. Indeed, very few algorithms have been applied to such graphs, and those that have, often take hours to run. For example, it is already very expensive to compute shortest distances from a single source, not to mention graph pattern matching by subgraph isomorphism, which is intractable in nature. Therefore, it is natural to ask whether parallel graph algorithms can scale to the largest publicly-available graphs using a single machine with a terabyte of RAM, processing them in minutes.

Computer Aided Design
Most basic algorithms have an immediate use in the domain of CAD, and optimizing them would bring immediate benefits. However, in practical cases, most of the algorithms would need to be re-designed to consider the specific characteristics of the problem and of the hardware device used to run them. For instance, optimizing movements for a flying-probe tester requires an approximate algorithm able to provide good results in a very short time. The basic search algorithm could be redesigned using SIMD paradigm for exploiting GPGPU, while, at the same time, a looser thread-level parallelism could be used to access a database or calculate fail-back solutions.

Learning Agents
The loose concurrency of independent devices can be exploited to increase the overall efficiency of a system at no additional cost by assigning a learning task to a number of cooperating agents. Most practical applications of Evolutionary Computation are limited in practice by the computational requirements of evaluating each solution (the fitness function). Being able to distribute the whole process of learning could be beneficial both in reducing the workload required to a single agent and in decreasing the probability to be attracted by local optima. As the different agents would be loosely connected, the results could be produced in unpredictable instants or never be available. However, these limitations are not incompatible with a process that is inherently able to exploit random variations, such as an Evolutionary Algorithm. The resulting algorithms will be able to handle asynchronicity and unreliability, using results only when available. A first application of the technology would be the study of social networks.

The candidate will mostly work on different hardware architectures, varying from HPC (High-Performance Computing), to desktop, and to embedded systems, in C/C++, high-level architecture-independent language, CUDA/OpenCL, and Python.
Outline of work plan: The work plan is structured in three years, as the Ph.D. program.

Year 1
The Ph.D. student will improve his/her knowledge of writing parallel/concurrent software with advanced languages, such as C++, Python or CUDA, covering aspects not analyzed in standard curricula. The student will study the portability of standard algorithms on parallel architectures, he/she will analyze data types and storage strategies, CPU to GPGPU communications, and he/she will better understand the scalability of an algorithm. The student will start to focus on the topics to analyze in the following two years, and he/she will also follow most of the teaching activity required by the Ph.D. curriculum. Such a preliminary research activity will mainly target IEEE or ACM conference papers.

Year 2
The work will concentrate on the implementations of parallel graph-based algorithms with strong bounds such as connectivity, bi-connectivity, strongly connected components, maximal independent set, maximal matching, graph coloring, single-source shortest paths, minimum spanning forest, and approximate set cover. He/she will try to achieve good performance on graphs with billions of vertices and hundreds of billions of edges. Interdisciplinary aspects will also be considered. Research will start to target journal papers. Credits for the teaching activity will be finalized. The work will mainly target software written on HPC and desktop architectures.

Year 3
The activity carried forward during the second year will be consolidated, targeting more specific applications in the CAD and in the Learning Agent domain. For this analysis, we will also vary the considered platforms moving from HPC and desktop architectures to embedded devices, mobile phone GPUs, and application specific computers such as the one within the NVIDIA Drive PX series.
Expected target publications: The target publications will be the main conferences and journals related to parallel computing, algorithms and computations, and CAD. Given the nature of the project, interdisciplinary conferences and journals will also be considered.
Thus, the following represent some of the admissible targets:
- Elsevier Software: Practice and Experience
- Elsevier Concurrency and Computation: Practice and Experience
- IEEE Transactions on Parallel and Distributed Systems
- IEEE Transaction of Intelligent Transportation Systems
- IEEE Transaction on Emerging Topics in Computing
- IEEE Access
Required skills and competences:
Current funded projects of the proposer related to the proposal:
Possibly involved industries/companies:Possibly, FCA and Magneti Marelli. SPEA is not involved in the proposal itself, but eventual results would be very interesting for them.

Title: Attention-guide cross domain visual geo-localization
Proposer: Barbara Caputo
Group website: https://scholar.google.com/citations?user=mHbdIAwAAAAJ
Summary of the proposal: Photo geolocation is a challenging task since many photos offer only few cues about their location. For instance, an image of a beach could be taken on many coasts across the world. Even when landmarks are present there can still be ambiguity: a photo of the Rialto Bridge could be taken either at its original location in Venice, or in Las Vegas. Traditional computer vision algorithms rely on the features provided to them during training. While this can lead to some degree of success when the data distribution of training and test data are the same, the problem becomes arduous when there is a domain shift between the two distributions. The goal of this PhD is to study the problem of visual geo-localization across visual domains. We will leverage over the intrinsic spatial connotation of place images and combine attention mechanisms with modern domain adaptation algorithms, in order to obtain perceptual representations that can be used for visual place recognition, as well as for content based image retrieval, able to close the domain gap differently on different parts of the images. Experiments will be conducted on publicly available databases as well as on data collections created during the project.
Rsearch objectives and methods: The problem of assigning a geo-localization label to an image has been extensively studied in the computer vision literature. The most important challenges in identifying places come from the complexity of the concepts to be recognized and from the variability of the conditions where the query images are captured. Scenes from the same category may differ significantly, while images corresponding to different places may look similar. Over the last years, deep learning based methods have become mainstream [a,b]. Some of these works showed the benefit of adopting a region-based approach (i.e. considering only specific image parts) in combination with descriptors derived from CNNs, so to obtain models robust to viewpoint changes and occlusions. Less researched is how to deal with shift in the distribution of the data used for training of the models with respect to those presented to the architecture at training time.The conceptual objective of this PhD project is to merge together research on attention driven visual geo-localization with unsupervised domain adaptation, with the goal to develop visual place classification algorithms able to deal with changes in the viewing conditions presented at training time like severe weather and illumination changes, or heavy occlusions like those caused by selfies. We will do so by leveraging over research on how to localize the spatial roots of domain shift in images [c, d]. Indeed, visual changes like those mentioned above have a clear spatial connotation, hence it is possible to localize and spatially ground the domain shift between source and target data. By doing so, we expect to be able to develop spatially localized domain adaptation architectures that will make it possible to recognize important landmarks in a scene even in presence of severe domain changes, or when a large part of a scene is occluded because of group photos.
The algorithms developed in the PhD project will be tested on publicly available databases, so to allow for a fair comparison with the current state of the art, as well as on a new database, to be acquired during the PhD, consisting of images of the most relevant Italian cities. The database will be initially created by leveraging over publicly available resources (like Google Street View), and then will be integrated with images acquired manually so to represent the most extreme cases of domain shift described above. The collection of the data described, that will be annotated and assessed, with the aim to become publicly available, will be the second objective of the PhD.

[a] T. Weyand, I. Kostrikov, J. Philbin. Planet-photo geolocalization with convolutional neural networks. Proc. ECCV 2016.
[b] M. Mancini, S. Rota-Bulo’, E. Ricci, B. Caputo. Learnind deep NBNN representations for robust place categorization. IEEE Robots, and Automation Letters, 2017.
[c] T. Tommasi, M. Lanzi, P. Russo, B. Caputo. Learning the roobts of visual domain shift. Proc. ECCV16
[d] G. Angeletti, B. Caputo, T. Tommasi. Adaptive deep learning through visual domain localization. Proc. ICRA18
Outline of work plan: The research workplan will be articulated as follows:

M1-M6: Implementation and testing of [a,b,c,d] on reference benchmark databases in conjunction with off-the shelf state of the art in visual place geo-localization; acquisition and annotation of a first version of a database containing images of Italian cities (Turin, Milan, Genoa, Rome etc) using Google Street View; implementation and testing of relevant baselines in the literature.

M7-M12: Design and implementation of a cross-domain local loss for adaptive geo-localization. Assessment of work on established benchmarks and comparison with previous work. Assessment of work on first version of the database of Italian cities. Writing of scientific report on findings of Y1.

M13-M24: Design and implementation of a cross-domain local loss for geo-localization robust to extreme domain shifts like selfies. Extension of the database of Italian cities to contain several images of selfies taken in such cities. Assessment of work on the established benchmarks and on the acquired database. Writing of scientific report on findings of Y2.

M25-M36: Implementation of final deep architecture for cross domain geo-localization incorporating the best results from Y1 and Y2, with principled reduction of the internal parameters of the network. Completion of the data acquisition for the database of Italian cities. Assessment of work on the established benchmarks and the new database. Writing of scientific report on overall findings of the project. Writing of PhD thesis.
Expected target publications: It is expected that the scientific results of the project will be reported in the top conference in the field of computer vision (IEEE CVPR, IEEE ICCV, ECCV). At least one journal publication is expected on one of the following international journals: IEEE PAMI, IJCV, CVIU.
Required skills and competences: The successful candidate should have a degree in computer science, computer engineering, mathematics or physics, with previous experience in machine and deep learning. A master thesis work on machine/deep learning applied to visual data, and previous publications in research topics related to the project are desirable.
Current funded projects of the proposer related to the proposal: CINI-VIDESEC (3Y, 1.2M €)
Possibly involved industries/companies:PCDM -DIS

Title: Analysis, search, development and reuse of smart contracts
Proposer: Valentina Gatteschi, Fabrizio Lamberti
Group website: http://grains.polito.it/
Summary of the proposal: In recent years, Blockchain technology received an increasing interest from researchers and companies. According to statistics, the global blockchain market is expected to grow from $1.2 billion in 2018 to more than $20 billion by 2024.
A key advantage of blockchain is that it can run smart contracts, (small) programs that are automatically executed (and that could trigger cryptocurrency transfers) when some conditions occur. Once a smart contract's code is stored on the blockchain, everyone can inspect it and it becomes no longer modifiable.
Despite the statistics, a considerable portion of the wider public still do not trust blockchain technology and smart contracts, mainly due to its complexity. In fact, understanding the behavior of a smart contract requires some (good) programming skills.
Another drawback of the current situation is that a comprehensive, easily-browsable repository to retrieve smart contracts’ code, in order to ease the development of new smart contracts is missing.
This proposal aims at addressing the above limitations by investigating new techniques and proposing novel approaches for the analysis, search, development and reuse of smart contracts. The results that will be achieved could be relevant for both developers, that could find/reuse smart contracts already developed in a given context, both people without technical skills, that could be able to understand the behavior of (or even code) a smart contract thanks to visualization and semantic technologies, among others.
Rsearch objectives and methods: A blockchain is a public ledger distributed over a network, recording transactions (messages sent from one network node to another) executed among network participants. Before insertion, each transaction is verified by network nodes according to a majority consensus mechanism. Recorded information cannot be erased and, at whatever time, the history of each transaction can be recreated.
Smart contracts are pieces of code, stored on the blockchain, programmed to behave in a given manner when certain conditions are met. They can be executed automatically without control of a third party.
Smart contracts have the following characteristics:
- they can trigger cryptocurrency transfers;
- their source code or their behavior can be inspected by everyone;
- they are not “legal” contracts, but are able to enforce payments;
- in case of malfunctions their actions cannot generally be reverted.

In order to develop smart contracts, programming skills are required. Some existing solutions propose template-based approaches to let non-skilled users deploy simple smart contracts on the blockchain. Nonetheless, this solution only works on a limited number of (simpler) cases. Similarly, only people with programming skills are able to verify if a smart contract’s code meets what is advertised. This limitation threatens the adoption of smart contracts by the wider public.

The activities carried out in this Ph.D. programme will aim at addressing the above limitations by investigating existing approaches, devising, and testing novel ones for:

a) the analysis of smart contracts’ source code/OPCODE: for what it concerns this objective, smart contracts’ source code/OPCODE deployed on the mainnet and on the testnet will be analyzed by considering different aspects, e.g., the quality of the written code, the behavior of the smart contract, the amount of triggered transactions, its context/objective, etc. The outcome of this activity will provide a methodology (or a tool) that could support developers and non-skilled people at a macro- and a micro-level. At a macro-level it could help them to understand how smart contracts are developed/used (or have been developed/used in the past), or to detect trends. At a micro-level, it could support them in quickly understanding the behavior of a smart contract. The analysis will be performed on the Ethereum blockchain, and eventually on other blockchains, to detect, for example, whether a given blockchain is preferred by a given sector, or to accomplish some specific tasks. Visual Analytics tools could be also exploited/developed to support the analysis.

b) the search of already developed smart contracts: to achieve this objective, a framework to categorize smart contracts analyzed in the previous phase will be designed. A methodology (or a tool) to search and retrieve the code of previously deployed smart contracts that could fulfill a given need will be devised. Semantics and Visual Analytics will be also used to support the search.

c) the development and reuse of smart contracts’ code: to achieve this objective, existing approaches for visual and natural language programming will be investigated to study their applicability to smart contracts. The result of this phase will be a methodology (or a tool) to support both programmers and non-skilled people in coding smart contracts from scratch or from existing ones (e.g., retrieved using the search engine devised in the previous phase).
Outline of work plan: The research work plan of the three-year Ph.D. programme is the following:

- First year: the candidate will perform an analysis of the state-of-the-art on available methodologies/tools for analysis, search and development/reuse of code, with a particular focus on smart contracts. The candidate will also deepen his/her competences on visual analytics, semantic technologies and natural language programming, among others. After the analysis, the candidate will identify advantages, disadvantages and limitations of existing solutions and will define approaches to overcome them.

- Second year: during the year, the candidate will design and develop methodologies and tools for analyzing smart contracts’ code, and will design a framework to categorize and search them.

- Third year: the third year will be devoted to the design and development of methodologies and tools enabling smart contracts reuse and coding, as well as to testing the developed tools.
Expected target publications: IEEE Transactions on Services Computing
IEEE Transactions on Visualization and Computer Graphics
IEEE Transactions on Knowledge and Data Engineering
IEEE Access
Future Generation Computer Systems
IEEE International Conference on Decentralized Applications and Infrastructures
IEEE International Conference on Blockchain and Cryptocurrency
IEEE International Conference on Blockchain
Required skills and competences: Good programming skills are required. Solidity programming skills are desired.
Current funded projects of the proposer related to the proposal:
Possibly involved industries/companies:

Title: Reliability and Safety of Artificial Neural Networks
Proposer: Matteo Sonza Reorda
Group website: http://www.cad.polito.it
Summary of the proposal: Artificial Neural Networks (ANNs) are increasingly used in many application domains, including some where safety is crucial (e.g., automotive and robotics). Possible faults affecting the hardware implementing the ANN can impact on the produced results. Unfortunately, we still miss a full understanding of which are the most critical faults, and which are the most sensible architectures (e.g., based on CPUs, GPUs, FPGAs, dedicated hardware). The goal of the proposed research activity is to fill this gap, leveraging on the available expertise and infrastructures developed by the CAD Group, e.g., in terms of GPU models and Fault Injection environments. A major issue to be faced is how to tame the computational effort required to consider the huge amount of faults that may affect the hardware, tracing their effects up to the application level. The research activity will benefit of the reliability analysis performed in the first phase to also develop solutions to increase the achieved reliability and safety, acting both on the hardware and software.
Rsearch objectives and methods: The planned research activities aim first at exploring the effects of faults affecting the hardware implementing ANNs, with special attention to GPUs. Experiments will study the effects of the considered faults on the results produced by the ANN. This study will mainly be performed resorting to fault injection experiments. In order to keep the computational effort reasonable, different solutions will be considered, ranging from simulation- and emulation-based fault injection up to multi-level one. The trade-off between the accuracy of the results and the required computational effort will also be evaluated.
Based on the results coming from the first phase, the second phase of the research activity will aim at devising solutions able to harden the ANN implementation with respect to hardware faults, acting either on the hardware or on the software. The impact of the proposed hardening solutions in terms of area and performance overhead will be evaluated.
Outline of work plan: The proposed plan of activities is organized in the following phases:
- phase 1: the student will first study the state of the art and the literature in the area of ANNs, their implementation on GPUs and their applications. Suitable cases of study will also be identified, whose reliability and safety could be analyzed with respect to faults affecting the underlying hardware. Suitable solutions to analyze the impact of faults will also be evaluated and put in place.
- phase 2: a set of fault injection campaigns to assess the reliability and safety of the selected cases of study will be performed, and the results analyzed.
- phase 3: based on the results of the previous phases, new techniques for ANN hardening will be devised and evaluated.
Phases 2 and 3 will also include dissemination activities, based on writing papers and presenting them at conferences. We also plan for a strong cooperation with the researchers of the Federal University of Rio Grande do Sul (Brazil), having special expertise in reliability evaluation, and with NVIDIA engineers.
This proposal is complementary and strictly related to the one titled "Autonomous systems' reliability and safety": both focus on the reliability of AI architectures for safety-critical embedded applications, but the two proposals deal with different underlying hardware architectures.
Expected target publications: Papers at the main international conferences related to test, reliability and safety (e.g., ETS, ATS, VTS, IOLTS, ITC).
Papers on the main IEEE journals related to design, test and reliability (e.g., IEEE Design and Test, IEEE Transactions on VLSI, IEEE Transactions on CAD, IEEE Transactions on Reliability, IEEE Transactions on Computers)
Required skills and competences: Basic skills in electronic circuit design.
Current funded projects of the proposer related to the proposal:
Possibly involved industries/companies:NVIDIA

Title: Autonomous systems' reliability and safety
Proposer: Ernesto Sanchez
Group website: http://www.cad.polito.it/
Summary of the proposal: Autonomous systems (ASs) have been a subject of great interest during the last years. In fact, autonomous devices have been investigated and proposed by very different actors: from the automotive industry to hospital UV disinfection robots; from automatic stacking systems to unmanned aerial vehicles. In order to support the software complexity, ASs may include some microprocessor cores, different memory cores, and hardware accelerators in charge to perform the Deep Artificial Neural Network computations.
An emerging problem is the verification, testing, reliability, and fault-tolerance of Autonomous Systems, and in particular, regarding the computational elements involved in the artificial intelligence computations.
These very complex systems are still lacking by a holistic analysis and comprehension related to reliability and safety. It is for example unclear how to functionally verify the behavior of a DNN, or what may happen when the autonomous system is affected by a fault from both points of view: reliability and safety.
During this project, the Ph.D. candidate will study from the hardware and software point of view, how to improve the reliability and safety of autonomous systems based on AI solutions.
Rsearch objectives and methods: The starting activity aims at studying the current design and verification methodologies that try to guarantee a correct implementation of AI based systems in ASs with particular interest on the available solutions to increase the systems safety.
During this phase, a set of benchmarks that will provide the suitable cases of study for the following research steps are defined. Two main types of AI systems in ASs will be analyzed: using hardware accelerators based for example in FPGA implementations, and supported by components-off-the-shelfs (COTS) such as systems embedding a high performance processor core.
Regarding the testing issues, it is necessary to define a set of fault models supported by a failure analysis of hardware implementations of the different systems and in particular to the ones related to the Neural Network implementations.
From the reliability point of view, there is a lack of metrics able to correctly assess how reliable is an AI-based system, in fact, a study and proposal of appropriate metrics is also required at this point.
Finally, mitigation strategies based on self-test and error-recovery mechanisms will be developed for the autonomous systems studied. The final goal is to equip the AI hardware with self-test mechanisms to detect hardware errors and fault-tolerance mechanisms for recovering from an error that has occurred and, thereby, continue the AI algorithm uninterruptedly maintaining the system accuracy.
Outline of work plan: Proposed work plan:

1. Study and identification of the most important works on design and verification of AI solutions used in autonomous systems.
2. Design and implementation of the cases of study resorting to hardware accelerators and COTS based on high performance processor cores.
3. Fault model definition and experimentation mainly resorting to the implemented cases of study. 4. Reliability and safety metrics definition.
5. Mitigation strategies proposal.
The first two steps will left the Ph.D. candidate with the appropriate background to perform the following activities.
Steps 3 to 5 are particularly interesting form the research point of view, allowing the student to write papers and present them in the International conferences related to the research area.
During these research phases, the student will have the possibility to cooperate with international companies such as NVIDIA and STMicroelectronics.
This proposal is complementary and strictly related to the one titled "Reliability and Safety of Artificial Neural Networks": both focus on the reliability of AI architectures for safety-critical embedded applications, but the two proposals deal with different underlying hardware architectures.
Expected target publications: The main conferences where the Ph.D. student will possibly publish her/his works are:
DATE: IEEE - Design, Automation and Test in Europe Conference
ETS: IEEE European Test Symposium
ITC: IEEE International Test Conference
VTS: IEEE VLSI Test Symposium
IOLTS: IEEE International Symposium on On-Line Testing and Robust System Design
MTV: IEEE International workshop on Microprocessor/SoC Test, Security and Verification
Additionally, the project research may be published in relevant international journals, such as: TCAD, TVLSI, ToC.
Required skills and competences:
Current funded projects of the proposer related to the proposal: N/A
Possibly involved industries/companies:NVIDIA, STMicrolectronics

Title: Orchestrating NFV edge services in a cloud-native world
Proposer: Fulvio Risso
Group website: http://netgroup.polito.it
Summary of the proposal: Network operators and telcos are increasingly relying on softwarized network functions (NFs) to replace dedicated network appliances. While bringing evident advantages in terms of agility, the drawback is in the performance of the current software, which sits well below the corresponding dedicated appliances counterpart.
The current PhD is oriented to investigate the above problem, focusing on telco-relevant (e.g., 5G) network services, with the ambitious objective of reaching terabit speed on common off-the-shelf hardware (e.g., commodity server), which corresponds to an improvement of one order of magnitude compared to current state of the art.
Rsearch objectives and methods: The candidate could pursue one (or more) of the following directions to reach the objective, while maintaining the "agile" properties of the current software models and the capability to run the software on common virtualization platforms (no dedicated servers):
- Exploit recent hardware accelerators such as SmartNICs or GPU cards;
- Novel hardware-aware algorithms (e.g., to consider the CPU cache structure) that can improve the efficiency of current hardware pipelines (e.g., CPU);
- Novel software techniques that can optimize the executed program at run-time, exploiting information such as the current configuration (e.g., rules configured in a firewall) and the current traffic (e.g., which rules are hit most frequently with the current traffic), while reducing the overhead across multiple network functions (e.g., header parsing shared across multiple network functions);
- Novel parallelization approaches that can allow to scale horizontally (multiple CPU cores exploited in parallel) and vertically (split the software across multiple software modules executed in sequence, on different servers).

Finally, a use case will be defined to validate the developed algorithms in realistic conditions. Among the possible choices, a 5G mobile gateway serving a set of customers, running on a server cluster with non-dedicated servers (e.g., the cluster can deliver both NFV and other generic services at the same time).
Outline of work plan: The proposed research plan is structured as follows.

Phase I (first year)
- State of the art
* Hardware accelerators
* Compilers and optimizations for data plane applications)
- Paper writing: survey and state-of-the-art of high-speed data plane
- Definition of single-program optimization algorithms
- Validation of the above algorithms in the Polycube eBPF framework
- Paper writing: novel compilation optimizations for data plane programs

Phase II (second year)
- Definition of cross-program optimization algorithms
- Validation of the above algorithms in the Polycube eBPF framework
- Paper writing: novel compilation optimizations for multi-stage data plane programs
- Introducing hardware-based (e.g., SmartNICs) optimization algorithms
- Validation of the above algorithms in the Polycube eBPF framework
- Paper writing: hardware-assisted algorithms for data plane processing software

Phase III (third year)
- Going datacenter-wide: co-existence of NFV tasks with traditional jobs: how to maximize the utility of the available cluster, how to provide strong guarantees (e.g., latency, throughput) to running services
- Validation of the above algorithms in the Polycube eBPF framework within the Kubernetes environment
- Paper writing: joint orchestration of data plane programs and general-purpose tasks on shared datacenters
- Writing PhD dissertation
Expected target publications: Top conferences:
- USENIX Symposium on Operating Systems Design and Implementation (OSDI)
- USENIX Symposium on Networked Systems Design and Implementation (NSDI)
- USENIX Annual Technical Conference (USENIX)
- International Conference on Computer Communications (INFOCOM)

Journals:
- IEEE/ACM Transactions on Networking
- IEEE Transactions on Computers
- ACM Transactions on Computer Systems (TOCS)

Magazines:
- IEEE Computer
- IEEE Networks
Required skills and competences:
Current funded projects of the proposer related to the proposal: None
Possibly involved industries/companies:TIM

Title: eXplainable Artificial Intelligence techniques for Natural Language Processing tasks
Proposer: Tania Cerquitelli, Elena Baralis
Group website: http://dbdmg.polito.it
Summary of the proposal: Despite the high accuracy promised by state-of-the-art deep learning models to address classification tasks, their applicability in real-life settings is still limited due to their opaqueness, i.e. they behave as a black-box to the end-user. The eXplainable Artificial Intelligence (xAI) field of research is seeking for new solutions that try to fulfill the existing gap between accuracy and interpretability, encountering many obstacles. The explainability of complex machine learning models applied to domains such as structured data and image classification has been currently explored and scientific milestones have been reached. However, the Natural Language Processing (NLP) is still lacking robust and specialized solutions.
The main research goal of this proposal is designing innovative xAI solutions that offer a level of transparency greater than existing methods tailoring to NLP. The vast majority of existing data algorithms are opaque – that is, the internal algorithmic mechanics are not transparent, in that, they produce output without making clear how they have done it. Innovative approaches will be devised to make the NLP algorithms human-readable and usable by both analysts and end-users to significantly increase explainability and user control in classification tasks.
Rsearch objectives and methods: The research objectives address the following issues:

xAI solutions for NLP tasks
The huge amount of data collected from people's daily lives (e.g. web searches, social networks, e-commerce) are textual data. Black-box predictive model tailored to NLP tasks increases the risk of inheriting human prejudices, racism, gender discrimination and other forms of bias.
As NLP algorithms increasingly support different aspects of our life, they may be misused and unknowingly support bias and discrimination, if they are opaque (i.e., the internal algorithmic mechanics are not transparent in that they produce output without making it clear how they have done so). Innovative xAI solutions, tailored to NLP tasks, will be designed to produce more credible and reliable information and services. They will play a key role in a large variety of application domains by making the results of the data analysis process and its models widely accessible.

Concept-drift detection for xAI solutions
When dealing with large data collections or complex textual datasets, the model trained in the past may be no longer valid. The identification of concept-drift can potentially become a key issue in the data analytics pipeline.
Adaptive and interpretable strategies will be studied and tailored to NLP tasks to avoid the expensive and resource-consuming procedure of model re-training when not necessary, and to understand why and how data have changed over time.

Data and Knowledge visualization.
Visualization techniques help humans to correctly interpret, interact, and exploit data and its value. Innovative visualization representations will be studied to enhance the interpretability of the internal algorithm mechanics. Keeping the user in the analytics loop by leveraging human visual perceptions on intermediate data analytics steps is an efficient and interpretable way to understand algorithms decisions.
Outline of work plan: During the 1st year, the candidate will study available solutions to enhance the transparency of data analytics algorithms and design innovative methods to allow users to control their personal data. New data collection methods will be defined to maintain users informed about the data collection process. The first attempt towards interpretability black-box predictive models will be tailoring the state-of-the-art solutions in the context of image classification to NLP tasks.

During the 2nd year, the candidate will study and define interpretable and innovative algorithms for NLP tasks by providing more credible and reliable explainability services and by making the results of the data analysis process and its models widely accessible. Selecting specific subsets from which interesting knowledge can be independently derived is of paramount importance to detect and understand data drifts. Visualization techniques will leverage the innovative xAI solutions by keeping the user in the analytics loop.

During the 3rd year, the candidate will design a smart user interface to effectively show the explainability of black-box models through the proposed solutions.
During the 2nd-3rd year, the candidate will assess the proposed solutions in diverse NLP tasks (e.g., sentiment analysis, topic detection).

During all three years, the candidate will have the opportunity to cooperate in the development of solutions applied to the research project and to participate to conferences presenting results of the research activity.
Expected target publications: Any of the following journals
IEEE TKDE (Trans. on Knowledge and Data Engineering)
ACM TKDD (Trans. on Knowledge Discovery in Data)
ACM TOIS (Trans. on Information Systems)
ACM TOIT (Trans. on Internet Technology)
Information sciences (Elsevier)
Expert systems with Applications (Elsevier)
Engineering Applications of Artificial Intelligence (Elsevier)
Journal of Big Data (Springer)

IEEE/ACM International Conferences
Required skills and competences:
Current funded projects of the proposer related to the proposal:
Possibly involved industries/companies:None

Title: Grounded Language Learning for Multimodal Understanding
Proposer: Barbara Caputo, Giuseppe Rizzo
Group website: https://www.iit.it/research/lines/visual-and-multimodal-applied-l...
Summary of the proposal: Understanding natural languages has become in the last years a real challenge in artificial intelligence. Thanks to worldwide research efforts towards the modelization of human language through deep learning and semantic methodologies, the modern algorithms are able to extract correlations among words in a huge collection of documents, thus defining the meaning of a word in terms of the surrounding context.

While this approach has led to great improvements in standardized benchmarks, the primary aim of the language is still missing. Children learn language as the necessity, inside a group, to refer to real world entities located around them using a shared symbolism, thus giving a grounding to words they speak.

This aspect of the language is still missing from modern algorithms involved in conversational scenarios often ending in a limited and sometimes poor conversational experience, due to the constraints that the user has to understand before starting the conversation, thus limiting its naturalness. Recent studies have started to address this problem under the research topic of Grounded Language Learning and Understanding.

We propose a research investigation to improve the capabilities of artificial agents to understand grounded language in order to correctly refer natural language sentences to visual scenes.
Rsearch objectives and methods: The investigation moves around two main topics: grounded language learning (Objective 1) and understanding (Objective 2-3). In parallel experiments on a test bed will be done in order to assess the real capabilities of such systems.

Objective 1: Learning
Design of a deep learning approach to align text captions to the referenced regions in a visual scene. State-of-the-art approaches propose transformer-based architecture with a co-attention mechanism to correctly mix the visual and textual information’s streams. We will investigate such architectures, understand their limits and problematics and try new approaches such as new flavors of attentions a new paradigm like self-supervised learning that can lead to performance improvements under three aspects:
1- Training time
2- Data needed
3- Final F1 performances
We aim to improve the current methodologies by using existing English dataset including both images and related textual information.

Objective 2: Understanding assessment
Design of a set of experiments to assess the understanding capabilities of these learning systems in real world problems. Research communities have just started to work on problems like visual dialogues involving an user, an agent and a real scene that can be an image or a video. These problems, along with visual question answering or egocentric vision, need to correctly understand the relationships between natural language sentences and the corresponding referred visual entity/object. These types of tasks involve both grounded learning and simulation of artificial reasoning that are hardly achieved in today's deep learning.

Objective 3: Understanding improvement
Design of novel architectures that can exploit both deep learning and semantic methodologies to address some open problems in understanding of grounded human commands. Reasoning problems emerge in grounded language understanding every time a sentence refers to position, attributes and relationships among objects in a scene and imply a downstream action to be performed. These problems are an ever-green of artificial intelligence research since the Winograd experiment with the “small blocks” world where an agent has to perform operation on colored blocks based on grounded commands. Today the research community has started to address such tasks with the use of Neuro Symbolic techniques consisting in the joint work between artificial neural networks and semantic methodologies such as ontologies and knowledge graphs.
Outline of work plan: Year 1 (Objective 1, Objective 2): Systematic literature review of the domain and an investigation of various standard benchmark datasets containing English text and referred images in order to train a model to learn the alignment between text and visual features.

Year 2 (Objective 2, Objective 3): Understand the utility of the developed technologies in downstream tasks involving a higher level of complexity due to the different context in which they live. The work will consist in the application of algorithms on tasks involving two modalities: vision and language. The tasks we will focus on are visual question answering and visual dialog. A core activity during this phase will be the analysis of the limits of current technologies in addressing particular tasks involving reasoning-like processes.

Year 3 (Objective 3): The third phase will be a natural continuation of the second one since it will cover the limits discovered in the current technologies. In order to achieve an “advanced” understanding we will investigate neuro symbolic methodologies trying to show their strengths and limits. During this phase we will focus on general question answering (GQA) and CLEVR datasets by Stanford.

During these three years we will continuously run validation experiments utilizing standard benchmark datasets in order to assess the real capability of the developed techniques.

The candidate will also be involved in research projects at LINKS Foundation and there will be the possibility to spend 1 year abroad in a top-tier research center.
Expected target publications: 3 top-tier conference papers such as EMNLP, NIPS, AAAI, ACL, CVPR, LREC
1 journal such as IEEE Transactions on Neural Networks and Learning Systems, Journal of Machine Learning Research (JMLR), Artificial Intelligence, Journal of Artificial Intelligence Research (JAIR), Artificial Intelligence Review
Required skills and competences: Artificial Intelligence
Machine Learning
Deep Learning
Computer programming (especially Python)
Current funded projects of the proposer related to the proposal: OBLIVION (H2020 EIT Digital no. 20319)
easyRights (H2020 DT-MIGRATION-06-2018-2019 no. 870980 )
Possibly involved industries/companies: LINKS Foundation

Title: Semantic image segmentation in open set scenarios
Proposer: Barbara Caputo, Fabrizio Dominici (LINKS)
Group website:
Summary of the proposal: Thanks to its wide range of applications, image analysis and in particular semantic image segmentation have recently become a central topic in computer vision research, where numerous studies have been published with remarkable results.
However, the vast majority of current techniques for image segmentation are based on supervised deep learning approaches and rely on a closed set of fixed categories. The supervised approach requires a huge amount of labelled data, imposing a strong limitation for real-world applications, where the annotations can be hard to obtain, or the required output categories may change in time. Nevertheless, a robust segmentation system should be able to provide delineation masks, even in case of unknown classes.
Therefore, new methods and approaches will be researched with a view to overcoming current limitations of image segmentation techniques, including but not limited to: the reduction of training data required to perform detection and segmentation tasks, the identification and the segmentation of unknown objects at test time, the contextual description of the image, the linking of open data and heterogeneous data sources.
Rsearch objectives and methods: Subdividing images into semantically comparable regions has become a crucial visual task that has recently received a lot of attention, thanks to its wide applicability in different domains, such as satellite monitoring or automated driving systems. Nowadays, state-of-the-art solutions typically include deep learning models with constantly increasing performance, slowly tightening the gap between human and machine performance.
Together with more powerful hardware becoming more widespread, a key factor that determined the substantial improvements in this sector is the growing number of carefully annotated datasets.
However, deploying models on real-world domains, characterized by dynamic environments and continuously evolving situations, presents numerous difficulties. For instance, the same robustness observed on the training set cannot be guaranteed with different contexts, or there could be no useful results at all when the categories are not exactly matching the training ones and the model is forced to choose among a fixed set of outputs.
Moreover, the provision of domain-specific labeling is usually unfeasible in terms of time and money, especially in complex segmentation tasks.
Due to the aforementioned and other issues, a viable solution is needed in order to delineate not only the closed set of categories, but every object present in the scene. This problem of detecting unknown classes is known in literature as Open Set recognition (OS). While recent works tried to tackle the problem from the point of view of image classification and object detection, the overall task can still be considered an open problem.
Some incremental advances towards a more complete segmentation have already been made, but they generally consist of partially supervised approaches, where additional information, such as raw bounding boxes, is exploited as a starting point to provide a more refined mask of the output. In many situations where data is scarce, weekly supervised segmentation remains difficult and requires additional retrieval or training steps, while in some other domains semi-supervised models may not apply at all because of lack of supporting data.
A key long-term purpose of the research will be to overcome the limitations of current segmentation architectures that operate on outputs known at train time by exploiting existing knowledge in similar domains to provide delineation masks for every category in the scene, despite the absence of label. New methods and approaches will be researched, including but not limited to: the reduction of training data required to perform the detection and segmentation tasks, the identification and the segmentation of unknown object at test time, the contextual description of the image, the linking of open data and heterogeneous data sources.
These techniques will then be tested on existing publicly available datasets, in order to be compared with current closed-set state of the art, and realistic settings of automated image segmentation from geospatial imagery, where annotated material remains rarely available. A specific focus will be devoted on the practical aspects of the problem, including the requirements needed in terms of computing resources to deploy the architectures for an extended and sustained usage.
Outline of work plan: Year 1: Analysis of the state-of-the-art algorithms and approaches for semantic segmentation, with a specific focus on OS segmentation. Implementation of existing baselines in the literature and extensive benchmark evaluation on publicly available datasets in various domains to identify strengths and weaknesses of existing approaches. Writing of a scientific report.

Year 2: Definition of a new Open Set Semantic Segmentation deep architecture, leveraging over the best practices in open set domain adaptation in the context of object classification. Comparison of results obtained with the methods examined in Year 1. Writing of a scientific report.

Year 3: Extension of approach developed in Year 2 to an architecture able to include symbolic knowledge in the form of textual attributes, to better explain the results obtained in the open set scenario and improve usability of the approach. Comparison with other explainable semantic segmentation and open set classification algorithms. Writing of scientific report and PhD thesis.

During the whole duration of the PhD work, the trained model will be further deployed in real-word scenarios provided by H2020 projects, funded by the European Commission.
Expected target publications: It is expected that the scientific results of this research will be published in conferences in the field of computer vision (e.g. IEEE CVPR, IEEE ICCV, ECCV) or machine learning (e.g. ICML, ICRL, NeurIPS), as well as relevant journals (e.g. IEEE PAMI, IJCV, CVIU, JMLR).
Required skills and competences: Good analytical understanding and problem solving
Good programming skills
Previous experience with machine and deep learning
Attitude towards self-improvement and teamwork
Current funded projects of the proposer related to the proposal: FASTER, SHELTER, VITIGEOSS (GA 869565, start 09/2020), SAFERS (GA 869353] start 10/2020), together with others that may be funded within the duration of the PhD period
Possibly involved industries/companies:LINKS Foundation

Title: Cross-domain federated graph learning
Proposer: Barbara Caputo
Group website: https://scholar.google.com/citations?user=mHbdIAwAAAAJ
Summary of the proposal: The current large success of data driven AI is strongly linked to its ability to leverage over very large amount of training data. While in some specific industries collecting and sharing these data is feasible, for several others this represents a challenge and it is not realistic that data acquired by different companies would be shared and stored in a unique central computing facility. Moreover, growing concerns about privacy and data protection call for a new generation of machine learning algorithms, able to leverage over data collected by a large network of devices without the data being moved from them. This PhD will focus on federated learning, a recently introduced machine learning framework based on data sets that are distributed across multiple devices while preventing data leakage. We will study federated learning when the data collected presents an intrinsic graph structure, explicitly dealing with the domain shift in the data caused by the distirbuted nature of the problem. While the application domain of the work will be in visual recognition and under standing, it is expected that results obtained in the thesis will be general and of interest for the whole machine learning community.
Rsearch objectives and methods: The current success of data-driven AI is largely due to Big Data availability, and as new data are continuously created and acquired by the growing spreading of sensing artificial devices. However, the real world situations are somewhat disappointing: with the exception of few industries, most fields have only limited data or poor quality data, making the realization of AI technology more difficult than we thought. Would it be possible to fuse the data together in a common site, by transporting the data across organizations? In fact, it is very difficult, if not impossible, in many situations to break the barriers between data sources. In general, the data required in any AI project involves multiple types. In most industries, data exists in the form of isolated islands. Due to industry competition, privacy security, and complicated administrative procedures, even data integration between different departments of the same company faces heavy resistance. It is almost impossible to integrate the data scattered around the country and institutions, or the cost is prohibited. At the same time, with the increasing awareness of large companies compromising on data security and user privacy, the emphasis on data privacy and security has become a worldwide major issue. The establishment of regulations like GDPR will clearly help build a more civil society, but will also pose new challenges to the data transaction procedures commonly used today in AI. To be more specific, traditional data processing models in AI often involves simple data transactions models, with one party collecting and transferring data to another party, and this other party will be responsible for cleaning and fusing the data. Finally a third party will take the integrated data and build models for still other parties to use. The models are usually the final products that are sold as a service. This traditional procedure face challenges with the above new data regulations and laws. How to legally solve the problem of data fragmentation and isolation is a major challenge for AI researchers and practitioners today.

This PhD thesis will investigate this learning framework within the context of visual recognition. The starting point will be federated learning [a], a new learning approach aiming to build models based on data sets that are distributed across multiple devices while preventing data leakage. We will explicitly take into account that devices acquiring data on a given visual recognition task are located in different visual domains, and hence the data recorded by each of them presents relative domain shifts. Moreover, we will address the fact that the network of devices can be modeled as a graph, and that graph-like structures might be present in the data acquired and should be properly modeled to solve successfully the task of interest. While the developed algorithms will be tested on publicly available visual databases, it is expected that results will be of interest for the whole machine learning community.

[a] Jakub Konecný, H. Brendan McMahan, Daniel Ramage, and Peter Richtárik. 2016. Federated Optimization: Distributed Machine Learning for On-Device Intelligence. arXiv:1610.02527
Outline of work plan: The research workplan will be articulated as follows:

M1-M6: Implementation and testing of state of the art federated learning algorithms on reference benchmark databases presenting domain shifts; implementation and testing of state of the art domain adaptation algorithms on state of the art domain adaptation reference benchmarks presenting a federated structure. Definition of a common reference scenario, comparison of existing methods.

M7-M12: Design and implementation of a cross-multi domain loss for federated domain adaptation. Assessment of work on the established benchmarks. Writing of scientific report on findings of Y1.

M13-M24: Design and implementation of a cross-multi domain loss for federated domain adaptation where the data domain presents a graph structure. Assessment of work on the on the established benchmarks. Writing of scientific report on findings of Y2.

M25-M36: Design and implementation of a cross-multi domain loss for federated domain adaptation where the data domain presents a graph structure, taking into account the graph nature of the federated learning problem. Assessment of work on the established benchmarks. Writing of scientific report on overall findings of the project. Writing of PhD thesis.
Expected target publications: It is expected that the scientific results of the project will be reported in the top conference in the field of computer vision (IEEE CVPR, IEEE ICCV, ECCV) and machine learning (ICML, NeurIPS, ICLR). At least one journal publication is expected on one of the following international journals: IEEE PAMI, IJCV, CVIU, JMLR, ML.
Required skills and competences: The successful candidate should have a degree in computer science, computer engineering, mathematics or physics, with previous experience in machine and deep learning. A master thesis work on machine/deep learning applied to visual data, and previous publications in research topics related to the project are desirable.
Current funded projects of the proposer related to the proposal: ICT48 ELISE
Possibly involved industries/companies:

Title: Crowd Monitoring in the Smart City
Proposer: Claudio Casetti
Group website: http://www.dauin.polito.it/research/research_groups/netgroup_comp...
Summary of the proposal: This proposal pertains to ongoing work that is being done in European Projects of the 5G-PPP calls, focusing mainly on upcoming 5G networks and their ecosystem, and it spans the domains of mobile communication, IoT systems and artificial intelligence. The overarching purpose of the work aims at the identification and quantification of people in sensitive areas (e.g., for safety and security purposes, such as during large crowd gatherings) or in areas of transit (e.g., for the purpose of dimensioning transportation networks or transit/parking/sheltering infrastructure, tourist attractions, etc.). While the detection of presence and head count is important, more valuable information would stem from the identification of flows of people. Cameras can be used for this purpose, although they require a high upfront investment, resource-consuming detection software, expensive maintenance, not to mention the privacy concerns they usually raise. Swifter, more lightweight solutions would instead focus on the use of sensors or radio scanners that capture the footprints of our mobile devices (e.g., the WiFi or Bluetooth beacons and probes they periodically transmit) and collect them (in strictly anonymous form). AI techniques can then be leveraged to automatically infer the presence and mobility patterns that result from the collection of such probes.
Rsearch objectives and methods: It is widely believed that IoT systems will have a momentous impact on people’s everyday lives, including leisure and touristic activities, as testified by the development of specific use cases for upcoming 5G networks. Nowhere will this impact be more tangible than in our “smart” cities, where density of network coverage, complemented by the deployment of sensors, can enable advanced applications. One of the key smart city scenarios revolves around the quantification of people in sensitive areas (e.g., for safety and security purposes, such as during large crowd gatherings) or in areas of transit (e.g., for the purpose of dimensioning transportation networks or transit/parking/sheltering infrastructure, etc.). A further step in this direction would stem from the identification of flows of people, e.g., transit rate in specific directions, or lingering times. Such information can provide site managers with the means to dimension areas and spaces inside the site, to design and organize the user experience, e.g., in a museum or in an art exhibition, and to monitor the formation of gatherings if social distancing is required.
Cameras can be used for this purpose, although they require a high upfront investment, resource-consuming detection software, maintenance, not to mention the privacy concerns they usually raise. Presence-monitoring passive sensors are not a reliable source of flow detection (they may be triggered by multiple passers-by without necessarily identifying their number, they cannot correlate the presence of the same visitors in different rooms etc.).
Alternative solutions exist, such as sensors that scan the ISM bands and passively capture probes transmitted by smart-phones as they try to identify known nearby WiFi Access Points (APs). Similar information can be derived by collecting Bluetooth probes. These sensors have some limitations: (i) they only detect people who carry a smartphone (although it can be argued that this is now the majority of passers- by); (ii) if used in a standalone mode, they only quantify the presence of people, not the path they are following; (iii) the information they expose is non-customizable and it is largely affected by implementation nuances in WiFi/Bluetooth probe timing, hence a considerable amount of inference is required. By scanning the ISM bands at 2.4 and 5 GHz, the scanners receive WiFi/Bluetooth probes transmitted by nearby smartphones and store information that can be extracted from such probes, namely the timestamp and the (hashed) MAC address of the sender. Probes are nominally transmitted by all mobile devices with a WiFi/Bluetooth interface and they serve the purpose of quickly identifying the presence of a nearby AP that they may have previously already associated to.
The use of AI technique would then be required in order to sift through the presumably vast amount of data collected by these sensors, whether they are placed in outdoor areas, or in indoor environments (such as on public transportation vehicles). The purpose of these techniques would then be to infer the presence and mobility patterns that result from the collection of such probes.
Outline of work plan: The first six months (M6) will be devoted to the establishment of the state-of-the-art in the field, also looking at practical experiences that may have been achieved in other smart city testbeds. Then, in the following 6 months, (M12) the work will be focused on the development of an IoT measurement collection platform, that integrates with existing IoT platforms such as OneM2M (https://www.onem2m.org). In this phase, existing sample commercial probe collectors such as the Libelium Meshlium (http://www.libelium.com/products/meshlium) will be evaluated and used for initial probe collection and analysis.

During the second year, until M24, the PhD candidate will work on the development of low-cost, easily deployable probe sensors that can be used in an integrated sensor network with the aim of deployment in a testbed environment, while further developing the analytic capabilities of the IoT measurement collection platform to include AI functionalities.

In the final year, until M36, the work will focus on the deployment and testing of the complete system in a smart city environment, possibly as a vertical application for existing 5G communication platform development within European or National projects.
Expected target publications: Conferences:
IEEE Infocom, IEEE GIoTS, IEEE CCNC, ACM Mobicom, ACM Mobihoc

Journals:
IEEE Transactions on Mobile Computing
IEEE Transactions on Networks and Service Management
IEEE Communication Magazine
Required skills and competences:
Current funded projects of the proposer related to the proposal: 5G-EVE
5Growth
Possibly involved industries/companies: