Empirical techniques and algorithms to develop a resilient non-supervised touch-based authentication system

Abstract

Touch dynamics (or touch based authentication) refers to a behavioral biometric for touchscreen devices wherein a user is authenticated based on his/her executed touch gestures. This work addresses two research topics. We first present a series of empirical techniques to detect habituation in the user’s touch profile, its detrimental effect on authentication accuracy and strategies to overcome these effects. Habituation here refers to changes in the user’s profile and/or noise within it due to the user’s familiarization with the device and software application. With respect to habituation, we show that habituation causes the user’s touch profile to evolve significantly and irrevocably over time even after the user is familiar with the device and software application. This phenomenon considerably degrades classifier accuracy. We demonstrate techniques that lower the error rate to 3.68% and sets the benchmark in this field for a realistic test setup. Finally, we quantify the benefits of vote-based reclassification of predicted class labels and show that this technique is vital for achieving high accuracy in realistic touch-based authentication systems. In the second half, we implement the first ever non-supervised classification algorithm in touch based continual authentication. This scheme incorporates clustering into the traditional supervised algorithm. We reduce the mis-classification rate by fusing supervised random forest algorithm and non-supervised clustering (either Bayesian learning or simple rule of combinations). Fusing with Bayesian clustering reduced the mis-classification rate by 50% while fusing with simple rule of combination reduced the mis-classification rate by as much as 59.5% averaged over all the users.