Keystroke based User Identification with XGBoost

Abstract

User identification or recognition based on keystrokes is the ability of the system to recognise or identify the persons or users who type the username, passphrase or password with a specific keystroke behaviour patterns. Therefore, user identification problem based on the keystroke behaviour patterns is posed as a multi-class classification problem, where some standard multi-class machine learning techniques identify multiple users. In this paper, we employed Extreme Gradient Boosting (XGBoost) technique for user identification in the context of behavioral biometrics using the keystroke dynamics features. Multiple users from CMU's keystroke dynamics dataset are analysed using Machine Learning (ML) Techniques such as the XGBoost, Multinomial Logistic Regression, Random Forest, Probabilistic Neural Network, Decision Tree, Multinomial Naive Bayes and a Multilayer Perceptron. The performance measure utilised for the user identification problem is accuracy. We observed that XGBoost produced the highest accuracy for user identification. We also performed statistical paired t-Test at 1% level of significance on the techniques employed for the user identification task. Therefore, our results were corroborated statistically too.