Behavioral biometrics for continuous authentication of smartphone users

Abstract

Smartphone devices have become essential for managing sensitive operations such as on-line banking, accessing medical records, making digital payments, and using government services. As a result, consumers have raised the demand for robust and user-friendly authentication systems. However, traditional smartphone authentication methods such as knowledge-based (e.g., PIN codes, passwords) and static biometric systems (e.g., finger-print, facial recognition) suffer from significant limitations. These methods are vulnerable to smudge and spoofing attacks, and fails to provide ongoing security once a device is un-locked. To address these challenges, we propose the design and evaluation of efficient, multi-modal continuous authentication systems leveraging behavioral biometrics. As a starting point, the research presents a comprehensive review of state-of-the-art meth ods for continuous authentication, highlighting recent progress in behavioral biometric modalities such as hand movement and touch gestures, and identifying limitations in existing datasets. Particularly the lack of synchronized multi-modal behavioral biometric modalities, combined with static biometric characteristics such as facial features. Following, an efficient continuous authentication system is designed, by investigating advanced feature selection to identify most relevant features, showing consistent improvement with the subset of selected features. Further a new continuous authentication system is introduced using self-supervised contrastive learning. The system employs a lightweight convolutional neural network architecture based on depthwise separable convolutions, achieving high accuracy verification and identification tasks while maintaining computational efficiency. By extending the original self-supervised contrastive learning framework introduced in the previous contribution, a multi-modal fusion framework is designed, by combining hand movement and touch gesture data. Fusion is performed at the feature-level, demonstrating robust performance even on small annotated datasets where labeled biometric data is scarce. Lastly, a novel multi-modal dataset, MM-BioSync, is introduced to address the lack of synchronized behavioral biometric public datasets. The dataset integrates data from frontfacing cameras, motion sensors, and touchscreen interactions. Experiments conducted on the dataset reveal that integrating all modalities captured during reading and writing tasks yields the best performance in user verification, underscoring the value of multimodal approaches for continuous authentication. The findings of this thesis highlight the potential of behavioral biometrics and multimodal fusion in enabling continuous user authentication. By addressing key challenges related to model performance, computational efficiency, and dataset availability, this work advances the state-of-the-art and provides a foundation for developing secure and user-friendly authentication solutions on smartphones.