PhD Programme in Computer Science and Mathematics

Pasquale De Marinis

Code: 38-033-02-DOT1302947-3792

CUP: H91I22000410007

Computer Vision techniques for sustainable AI applications using drones

The research topic will focus on the development of new Computer Vision and Deep Learning techniques for sustainable AI applications through drones, particularly in precision agriculture. Drones are increasingly being used as cost-effective and agile tools for data collection in this field. They can be equipped with sensors, such as multispectral sensors, to gather valuable information for monitoring purposes. These data can then be processed using Computer Vision and Deep Learning techniques to drive decision-making processes, potentially leading to environmental and economic benefits. The research will concentrate on developing methods that offer a good compromise between effectiveness and efficiency.

Co-funded by: Exprivia S.p.A.

Andrea Esposito

Code: 38-033-02-DOT1302947-3664

CUP: H91I22000410007

Human-Centered Artificial Intelligence (HCAI) Techniques for Supporting End Users Interacting with AI Systems

Artificial Intelligence (AI) is becoming ubiquitous. The strive for automation that characterizes typical AI-based systems flourishes on two faulty assumptions: all users have the same needs and the AI decision should be trusted. By relying on a black-box model trained on a dataset, thus providing a fully-automated system, humans’ safety is bound to be threatened. On the other hand, a system that is not automated at all also poses risks to humans’ safety. A better approach lies in the middle: systems should provide the right level of automation, while guaranteeing a suitable level of control to the user. In such cases, the system can be considered “reliable, safe, and trustworthy”. To yield systems that possess these three properties, the final users must be involved in all steps of the AI system design process, aiming at providing a high level of human control while still offering a high level of automation when desired. In this context, this Ph.D. project aims at: (i) Providing more precise definitions of some basic concepts like “trustworthiness”; (ii) Exploring customization options of AI models; (iii) Identifying best practices and design requirements for human-centered AI models.

Co-funded by: Eusoft S.r.l.

Daniela Grassi

Code: 38-033-02-DOT1302947-3436

CUP: H91I22000410007

Emotion and Stress Recognition in Work Environment Using Biometric and Behavioral Data

This research investigates the feasibility of recognizing emotions through noninvasive, low-cost biometric sensors. The focus is on identifying risks associated with negative emotions and detecting stress and fatigue. Objectives include: (1) Developing methods and tools to enhance developers' emotional awareness and wellbeing. This involves identifying fatigue, stress, and anxiety, linking them to their triggers, and implementing strategies to restore productivity. Additionally, biometric data can help understand the impact of remote/hybrid work on employee health and wellness. (2) Gaining insights from multi-modal emotion recognition techniques, combining approaches like self-reported emotions and sensor-based detection. This will assess the alignment between automatically captured emotions and those reported through cognitive processes.

Co-funded by: Exprivia S.p.A.

Francesco Greco

Code: 38-033-02-DOT1302947-3793

CUP: H91I22000410007

Investigating XAI techniques to help users defend against phishing attacks

Phishing is a major cyber threat in which criminals attempt to illegally obtain personal information, generally by sending malicious emails. Despite advanced technical security measures, phishing remains successful due to inherent human vulnerabilities such as users' lack of knowledge, stress, etc. Warning dialogs that are used to protect users often fail as they do not address these human factors, e.g., they lack contextual information about why an email is dangerous. This can lead to a lack of user trust, resulting in reduced security. To interpret the underlying AI systems used for phishing classification and provide such information, XAI (Explainable Artificial Intelligence) solutions can be used. However, explanations must be understandable also by lay users to properly warn them. This project aims to tackle the problem from both a technological and a human perspective, investigating XAI solutions to detect phishing content and design effective user interaction with human-centered approaches.

Co-funded by: Auriga S.p.A.

Raffaele Scaringi

Code: 38-033-02-DOT1302947-3669

CUP: H91I22000410007

Automatic analysis of artistic heritage via Artificial Intelligence

Cultural heritage, particularly the visual arts, is of inestimable importance for our societies’ cultural, historical, and economic growth. Thanks to technological improvements, there has been a large-scale digitization effort in recent years, which has led to the increasing availability of large digitized art collections. In addition, the rise of Deep Learning and Computer Vision has allowed the scientific community to develop new methods to study and analyze these visual data. This may produce new systems that can support art experts and users in general. To this end, in this project, the focus will be on investigating graph learning approaches, combined with advanced Computer Vision techniques, to deal with the complex task of analyzing digitized fine arts. The goal is to develop effective, efficient, and interpretable models that can support art experts in the study of fine arts.

Co-funded by: Exprivia S.p.A.

Gaetano Settembre

Code: 38-033-02-DOT1302947-3608

CUP: H91I22000410007

Low-rank models for the analysis of Earth Observation data focusing on coastal and marine environments

Recent advancements in remote sensing have led to the creation of vast and diverse datasets, necessitating their thorough analysis and integration to derive actionable insights. This integration involves merging varied data types from sources like satellite imagery, aerial photography, and ground measurements. A new study focuses on developing low-rank approximation (LRA) models to merge these datasets, including hyper/multispectral images, LiDAR, and ancillary data, to better understand environmental phenomena. While current LRA methods are powerful, they may not fully meet real-world needs or preserve intrinsic data properties. The project seeks to enhance LRA models by incorporating physics-informed variants, utilizing optimization algorithms and cost functions to impose specific constraints on the extracted latent components from Earth observation data. This approach aims to improve the quality of approximation models and enhance their applicability across various domains.

Co-funded by: Planetek Italia S.r.l.