One-Class Models for Continuous Authentication Based on Keystroke Dynamics

In this paper we discuss an applied problem of continuous user authentication based on keystroke dynamics. It is important for a user model to discover new intruders. That means we don't have the keystroke samples of such intruders on the training phase. It leads us to the necessity of using one-class models. In the paper we review some popular feature extraction, preprocessing and one-class classification methods for this problem. We propose a new approach to reduce dimensionality of a feature space based on two-sample Kolmogorov-Smirnov test and investigate how the quantile- based discretization technique can improve the one-class models' performance. We present two algorithms, which have not been used for keystroke dynamics before: Fuzzy kernel-based classifier and Random Forest Regression classifier. We conduct experimental evaluation of the proposed approach.