Privacy Preserving Loneliness Detection: A Federated Learning Approach

Abstract

Today's smartphones have sensors that enable monitoring and collecting data on users' daily activities, which may be converted into behavioral indicators of users' health and well-being. Although previous research has used passively sensed data through smartphones to identify users' mental health state, including loneliness, anxiety, depression, and even schizophrenia, the issue of user data privacy in this context has not been well addressed. Here we focus on the feeling of loneliness, which, if persistent, is associated with a number of negative health outcomes. While modern artificial intelligence technology, specifically machine learning, can assist in detecting loneliness or depression, current approaches have applied machine learning to centrally collected user data at a single location with the potential to compromise user data privacy. To address the issue of privacy, we investigated the feasibility of using federated learning on single user data to identify loneliness collected by different smartphone sensors. Federated learning can help protect user privacy by avoiding the transmission of sensitive data from mobile devices to a central server location. To evaluate the federated method's performance in detecting loneliness, we also trained models on all user data using a centralised machine learning approach and compared the results. The results indicate that federated learning has considerable promise for detecting loneliness in a binary classification problem while maintaining user data privacy.