Machine learning has been proven to provide good performances on stress detection tasks using multi-modal sensor data from a smartwatch. Generally, machine learning techniques need a sufficient amount of data to train a robust model. Thus, we need to collect data from several users and send them to a central server to feed the algorithm. However, the uploaded data may contain sensitive information that can jeopardize the user's privacy. Federated learning can tackle this challenge by enabling the model to be trained using data from all users without the user's data leaving the user's device. In this study, we implement federated learning-based stress detection and provide a comparative analysis between individual, centralized, and federated learning. The experiment was conducted on WESAD dataset by using Logistic Regression as the classifier. The experiment results show that in terms of accuracy, federated learning cannot reach the performance level of both individual and centralized learning. The individual learning strategy performs best with an average accuracy of 0.9998 and an average F1-measure of 0.9996.

1st Medical Faculty Charles University Prague 12800 Prague Czech Republic

Department of Information Security and Communication Technology Norwegian University of Science and Technology 2815 Gjøvik Norway

eHealth Competence Center Bavaria Deggendorf Institute of Technology 94469 Deggendorf Germany

Medical Faculty University of Regensburg 93053 Regensburg Germany

References provided by Crossref.org