A hybrid novelty score and its use in keystroke dynamics-based user authentication

The purpose of novelty detection is to detect (novel) patterns that are not generated by the identical distribution of the normal class. A distance-based novelty detector classifies a new data pattern as "novel" if its distance from "normal" patterns is large. It is intuitive, easy to implement, and fits naturally with incremental learning. Its performance is limited, however, because it relies only on distance. In this paper, we propose considering topological relations as well. We compare our proposed method with 13 other novelty detectors based on 21 benchmark data sets from two sources. We then apply our method to a real-world application in which incremental learning is necessary: keystroke dynamics-based user authentication. The experimental results are promising. Not only does our method improve the performance of distance-based novelty detectors, but it also outperforms the other non-distance-based algorithms. Our method also allows efficient model updates.