Enhancing Client Privacy in Physiology-Based Biometric Verification with Differential Privacy and Positive-Label Federated Learning
Mohamed Benouis, Bhargavi Mahesh, Elisabeth André, Yekta Said Can · 2024
In recent years, the widespread adoption of multimodal physiological signal-based biometric systems has led to a significant increase in data exchange on cloud servers. This surge in data exchange has raised concerns over data breaches and privacy violations. Eventually, classical privacy protection methods become increasingly complex and impractical in meeting the security and privacy requirements. A cloud-server-based concept of federated learning mitigates the privacy issue and has shown promising performance improvements. Nevertheless, federated learning does not explicitly limit the information leakage of user data. Although research recommends applying perturbation mechanisms to the global server's learning process, these techniques may not effectively withstand inversion attacks, and as a result, they may not prevent unauthorized access attempts. In this paper, we introduce a secure federated learning method that leverages differential privacy (DP) techniques to enhance the precision of biometric systems while safeguarding individual client privacy through explicit constraints on data used for training, and consequently sharing among clients. Moreover, the integration of differential privacy safeguards user data privacy while sharing the local model with the global server. To assess the effectiveness of this approach in the federated learning scenario, experiments are conducted on two different datasets: WESAD and DAPPER. The experimental results demonstrate that the proposed system performs well while preserving the privacy of the data and labels.