Evaluating Learning Algorithms for Keystroke Based User Authentication

The field of keystroke-based authentication increasingly relies on AI technologies for increased robustness and accuracy. A number of such approaches have been recently proposed, with variable levels of success and computational demands. This paper aims to investigate the comparative performance of supervised and unsupervised learning using two algorithms, KNN and K-means++, and explore the impact to the keystroke-based user authentication. Three keystroke features are selected: dwell time, flight time and press-to-press latency. FAR, FRR and accuracy are used as the performance metrics of KNN while purity and silhouette coefficient are selected as the performance metrics of K-means++. The experimental results show that KNN is more suitable for the analysed scenarios and has a slightly higher accuracy of 74.58% than K-means++. We further propose a method of reprocessing the dataset based on modifying the outliers when unsupervised algorithm was also able to get very good performance with 0.8767 of purity.