Deep Learning Application in Continuous Authentication

This paper reviews deep learning architectures used in continuous authentication, emphasizing on sensor-based or implicit device authentication methods such as motion patterns and keystroke dynamics. We review fundamental architectures (RNN, CNN, Attention Mechanism) and more complex architectures like autoencoders and Siamese networks, which can incorporate various architectural units and their applications for continuous authentication. We have discussed each architecture’s benefits, limitations, and potential, as well as an option for hybrid architecture use. Further, we review the applications of these architectures in biometric verification (e.g., ECG), behavioral measures (e.g., keystroke or mouse movement), and behavioral biometrics using IMU data (e.g., accelerometer or gyroscope). After we encapsulate our previous research and propose the end-to-end methodology for continuous authentication with autoencoder-based models. We discuss the challenges and future trends in utilizing deep learning for continuous authentication, considering sensitive data privacy concerns and the need for developing highly robust and secure applications. Overall the paper aims to provide a general view of the usage of deep learning for continuous authentication, highlighting future directions and providing a view of the most commonly used approaches in the field for researchers.