Keystroke Dynamics User Authentication Based on Gaussian Mixture Model and Deep Belief Nets

Figure 2

Keystroke dynamics features for static key string “.tie5Roanl” from the CMU keystroke dynamics benchmark dataset [11]. The dwell time and digraphs for the first four data-collection sessions for three subjects are shown. Although the keystroke features provide sufficient distinguishing patterns for each subject, they are highly correlated with large-scale variations and typical of noise and outliers. We have previously proposed a new distance metric to effectively handle these challenges that are intrinsic to keystroke dynamics data [12]. In this work, we show that GMM-UBM and DBN-based approaches perform even better to model large variations and correlations in the data.