Analysis of template update strategies for keystroke dynamics

Abstract

Keystroke dynamics is a behavioral biometrics showing a degradation of performance when used over time. This is due to the fact that the user improves his/her way of typing while using the system, therefore the test samples may be different from the initial template computed at an earlier stage. One way to bypass this problem is to use template update mechanisms. We propose in this work, new semi-supervised update mechanisms, inspired from known supervised ones. These schemes rely on the choice of two thresholds (an acceptance threshold and an update threshold) which are fixed manually depending on the performance of the system and the level of tolerance in possible inclusion of impostor data in the update template. We also propose a new evaluation scheme for update mechanisms, taking into account performance evolution over several time-sessions. Our results show an improvement of 50% in the supervised scheme and of 45% in the semi-supervised one with a configuration of the parameters chosen so that we do not accept many erroneous data.