Continuous learning: a feasible solution for continuous authentication using PPG?

In this paper, we review the latest experiments on continuous biometric authentication using PPG, mainly those carried out by Aly and Di Pietro [2]. In their latest model, they propose a Local Outlier Factor (LOF) [1] binder and continuous training to minimize legitimate user data for training a model that they claim to be capable of self-repair after attacks. Here we delve into the subject of LOF and implement various corruption attacks. With this preliminary study, we show that, on the contrary, the system is not capable of repairing itself. Moreover, continuous blind training is not very effective, and may be counter-productive in some cases. Our first results demonstrate the need for further research into continuous authentication and continuous learning. We show that current metrics and methods are not sufficient to properly assess the relevance of PPG-based continuous authentication systems.