Personalization of AI Models Based on Federated Learning for Driver Stress Monitoring

Houda Rafi, Yannick Benezeth, Philippe Reynaud, Emmanuel Arnoux, Fan Yang, Cédric Demonceaux

Abstract: To improve the comfort of car occupants or to develop control laws for autonomous vehicles or Advanced Driver-Assistance Systems, it is essential to monitor drivers’ internal state and automatically detect stressful situations. In this paper, we propose a driver’s stress monitoring system based on the analysis of physiological signals. To consider the individual differences between drivers, we propose a training strategy based on federated learning that favors examples in training set from drivers with the same profile as the driver we want to monitor. This approach allows us to personalize the prediction model for a target-driver and significantly improves performance compared to the classical paradigm that maximizes the average performance for all the users in a given dataset. This paper shows that this personalization strategy improves the performance of the stress estimation on the public database AffectiveROAD [1].