Beyond~Login: Self-Supervised Multimodal Continuous Authentication for Smartphones

Abstract

The widespread storage of sensitive personal and financial data on mobile devices exposes a critical security gap in traditional point-of-entry authentication. Conventional mechanisms, such as PINs or static biometric scans, provide only initial verification and fail to maintain security throughout a device session, leaving devices vulnerable to session hijacking or opportunistic theft. To address this limitation, we propose a Self-Supervised Continuous Authentication (SSCA) framework that provides non-intrusive, real-time user verification through a parallel two-phase security sieve. Our architecture leverages two complementary behavioral modalities. Gait dynamics are modeled via a long short-term memory autoencoder to capture latent temporal motion patterns, while touch dynamics are evaluated using a one-class support vector machine to detect deviations in screen interaction behavior. Unlike conventional supervised approaches, our self-supervised method learns only the authorized user’s behavior, eliminating the need for labeled intruder data and preserving user privacy.