Dynamic swipe gestures based Continuous Authentication using similarity learning

Abstract

Continuous authentication is increasingly becoming a necessity for cell phones. Numerous mobile authentication methods have been proposed, such as continuous facial recognition, continuous voice recognition, keystroke dynamics and motion dynamics. This research addresses the growing need for continuous authentication on mobile devices, proposing a novel solution centered around swipe gestures. Unlike conventional methods utilizing machine learning on discrete swipe data, our approach treats swipes as images, tapping into the power of deep learning to extract nuanced features. We enhance these images with additional information, using color, brightness, and thickness to represent swipe speed, pressure, and finger size, respectively. Departing from the classification approach of previous studies, we employ similarity learning to detect differences between swipe gestures without requiring readaptation for each user. Facing the challenge of limited mobile device resources, we adopt a lightweight architecture, MobileNetV2, and further optimize it using Neural Architecture Search (NAS). This tailored approach yields promising results, achieving 98.33% accuracy with NAS and 97.77% with MobileNetV2.