Transfer learning in long-text keystroke dynamics

Abstract

Conventional machine learning algorithms based on keystroke dynamics build a classifier from labeled data in one or more sessions but assume that the dataset at the time of verification exhibits the same distribution. Ideally, the keystroke data collected at a session is expected to be an invariant representation of an individual's behavioral biometrics. In real applications, however, the data is sensitive to several factors such as emotion, time of the day and keyboard layout. A user's typing characteristics may gradually change over time and space. Therefore, a traditional classifier may perform poorly on another dataset that is acquired under different environmental conditions. In this paper, we apply two transfer learning techniques on long-text data to update a classifier according to the changing environmental conditions with minimum amount of re-training. We show that by using adaptive techniques, it is possible to identify an individual at a different time by acquiring only a few samples from another session, and at the same time obtain up to 19% higher accuracy relative to the traditional classifiers. We make a comparative analysis among the proposed algorithms and report the results with and without the knowledge transfer. At the end, we conclude that adaptive classifiers exhibit a higher start by a good approximation and perform better than the classifiers trained from scratch.