Machine Learning Techniques for Keystroke Dynamics

Conventional security mechanisms such as token-based and knowledge-based authentication mechanisms are losing importance in the present era of immense technological development in cyber threats. Password and pin are examples of these mechanisms. Keystroke biometrics is a promising solution for ensuring cybersecurity in both standalone and connected systems. Keystroke biometrics is a subset of behavioral biometrics and distinguishes users based on their typing patterns. The performance of a user authentication system utilizing keystroke biometrics depends on the extracted features and classification techniques. The objective of this paper is to compare three different learning techniques namely support vector machine, random forest and logistic regression, in the context of keystroke biometrics. Time-based features are extracted from a publicly available dataset. These features are analyzed with above mentioned machine learning algorithms, and the performance of these algorithms is compared. Hyperparameter tuning and cross-validation are performed to further enhance the performance. Experimental results demonstrate that Random forest is the most efficient with accuracy of 0.85 and F1 score of 0.74. The accuracy obtained with support vector machine and logistic regression is 0.76 and 0.63, respectively.