Predicted Templates: Learning-curve Based Template Projection for Keystroke Dynamics

Abstract

Keystroke Dynamics (KD) as a biometric modality can provide authentication tools in many real-life applications, virtually at zero-cost on the client side, due to the reliance of these techniques on existing hardware, and their low computational expense. One promising application is the use of KD as a second factor in password-based authentication. A downside of the existing modeling methods is the assumption of stationary behavior from the clients. However, it is expected that humans show improvements in performing a specific task following practice. In this study, we propose methods for utilization of learning models in predicting the future behavior of the clients, even with little enrollment data, and generate predicted behavioral models that can be used in different classifiers. In our experiments, the predicted templates show a reduction in the average equal-error-rate (EER) consistently across different classifiers a benchmark dataset. A reduction of 20% is achieved on the best classifier. Given fewer enrollment data, the performance gain was shown to reach above 30%. Furthermore, we show that blind detection of attacks is possible, solely relying on the global learning curve, with an EER of 16%.