A continuous user authentication scheme for mobile devices

Face and touch modalities have independently been shown to yield promising results for continuous user authentication. In this study, we present a novel framework that combines these modalities. We show a stacked classifier approach can be used to improve the continuous authentication on mobile devices and address some prevalent issues with the current state-of-the-art. We use a state-of-the-art public dataset containing face and touch-gesture modalities for 50 users. Features are extracted from each modality for each user. We train a set of classifiers for user modalities to provide probability scores on a sample. The scores capture the nuances of each sample and are concatenated into a vector. This vector is used in a meta-level classifier. The scores we obtain from the meta-level classifiers show our approach performs better than previous continuous authentication approaches. We achieve an equal error rate of 3.77% for a single sample. We also show the added robustness a multi-modal approach provides if one modality is compromised.