Smartphone Continuous Authentication Using Deep Learning Autoencoders

Continuous authentication is receiving increased attention from providers of on-line services, particularly due to the ability of mobile apps to collect user-specific sensor data. However, the approaches proposed so far are either not accurate enough to provide a high-quality user experience or restricted by engineering challenges to capture data continuously. In this paper, we propose an approach based on a deep learning autoencoder, which achieves an equal error rate as low as 2:2% in tested real-world scenarios. The suggested system only relies on accelerometer data and does not require a high number of features, therefore reducing the computational burden. We discuss the balance between the number of dimensional features and the re-authentication time, which decreases as the number of dimensions increases. We also discuss parameter selection for real-world scenarios e.g. depth of the architecture, time elapsed before re-building the model and length of the training dataset and possible approaches to find the optimal trade-off between accuracy and usability required for each particular context.