Transformer-Based Biometrics Method for Smart Phone Continuous Authentication

As a widely used personal device, smartphones handle substantial amounts of sensitive information, rendering their security a critical concern. To ensure a seamless user experience and achieve robust continuous authentication, this work introduces a deep learning-based model leveraging a transformer architecture. The model captures complex temporal dependencies by learning user behavior patterns from keystroke dynamics and sensor data (e.g., accelerometer and gyroscope), significantly improving the precision and robustness of multimodal biometric recognition. A dynamic behavior-weighted module is incorporated to optimize the multi-head attention mechanism, while the integration of a quadruple loss function further enhances biometric recognition accuracy. Experimental results validate the model's efficacy, demonstrating high usability and a low error rate in multimodal biometric datasets, positioning it as a viable solution for continuous implicit authentication on smartphones.