L'apprentissage fédéré personnalisé pour l'évaluation des conducteurs à partir de signaux physiologiques dans les ADAS

Humans are continuously confronted with environmental stimuli, to which they autonomically react behaviorally, physiologically, and emotionally, depending on personal judgement. In highly demanding situations, the emotional reaction is expressed as stress, while the physiological response involves measurable changes such as variations in heart rate, and electrodermal activity. These physiological changes serve as reliable markers for assessing stress, providing the possibility of precise and continuous stress monitoring that extends beyond traditional self-reports. These advantages make physiological stress assessment valuable in the automotive domain, where it is used to assess the psychological impact and validation of Advanced Driving Assistance Systems (ADAS) for the driver, guiding the refinement of stressful features while promoting those that enhance comfort and safety. To implement a deployable solution meeting industrial and methodological requirements, it is essential to consider the inter-individual variation and sensitivity of physiological signals, as well as the subjective and difficult-to-annotate nature of stress, often resulting in scarce labelled datasets and the need to accommodate new unlabelled individuals. In this thesis, we develop personalized federated learning (PFL) frameworks that enhance the decentralized paradigm by accounting for individual variability while preserving data privacy. These frameworks leverage the similarities across drivers' physiological response patterns to guide personalized knowledge transfer, with the aim of characterizing more homogeneous physiological and psychological profiles. The first proposed DP-FL framework leverages similarity to weight client contributions, whereas the second, TPCFL, exploits it to form more homogeneous clusters and enable individualized model selection, effectively addressing non-IID data and yielding superior results. The improvements observed across driving datasets supported the personalization approach, later refined with Transformer models to capture fine-grained relationships even in small datasets, and further extended through sTPCFL to ensure scalability to large populations. A deployable and practical protocol for the labelling of stress in driving scenarios was also proposed and empirically tested, based on independent self-reports with minimal delay, thereby overcoming the limitations of existing driving physiological datasets that rely on external observers or retrospective annotations.