Contrastive Learning for Continuous Touch-Based Authentication

The frequent handling of sensitive information on mobile devices has created an urgent need for robust security measures. Touchscreens, as the primary medium for human-computer interaction, offer an ideal solution for non-intrusive security through continuous authentication based on touch behavior. To address the limitations of existing single-modal binary classification methods, this paper proposes a unified contrastive learning framework. The framework utilizes a Temporal Masked Autoencoder to extract temporal features from multi-sensor data streams. This framework further integrates a Siamese Temporal Attention Convolutional Network and attention mechanisms to jointly model sequential and cross-modal features. Experiments conducted across multiple datasets demonstrate that this method outperforms current mainstream solutions, thus offering a reliable and efficient solution for mobile user authentication.