User clustering based on keystroke dynamics

Abstract

The PAM clustering algorithm is applied on the Si6 keystroke dataset in order to identify sessions of the same users. A number of heuristical outlier lters based on statistical properties of keystroke latencies are proposed and run on the dataset. Di erent tests are performed varying the number of digraphs that compose each observation and its dimensionality, in order to verify the assumption that more data gives a better quality of clustering and to estimate the minimum required number of dimensions. The number of clusters is estimated through the silhouette algorithm. Resulting clustering accuracy is measured by means of the F-measure, showing the viability of user identi cation through keystroke analysis.