Federated Learning and Biometric Identification for Continuous User Authentication Using Hybrid Neural Models

Abstract

Continuous user authentication is one way to improve security, especially in environments where traditional password-based systems have outlived their effectiveness. A promising solution for user identification will be behavioral biometrics, which map internal patterns – such as keystrokes and mouse movements. Centralized storage of biometric data presents privacy issues. Toward this goal, we propose a federated learning approach for user identification and authentication that integrates a hybrid CNN-RNN model to efficiently capture spatio-temporal patterns of user behavior. While RNN captures sequence information, CNN focuses on feature extraction. This can be leveraged within Federated Learning so that biometric data remains on the user's device, significantly reducing privacy risks. Experimental results show that the method proposed in this work is highly accurate for user identification with very low exposure to his data, highlighting the benefits of Federated Learning, i.e. effective improvement in both security and privacy.