IMPROVING KEYSTROKE DYNAMICS AUTHENTICATION: BALANCING ACCURACY AND USER EXPERIENCE THROUGH EFFICIENT TRAINING

Abstract

The aim of this study is to enhance the security and user experience of two-factor authentication systems through the application of keystroke dynamics, a form of behavioral biometrics. Keystroke dynamics analyze the unique typing patterns of users to offer a biometric factor for authentication, which complements the traditional knowledgebased method (username and password). This study specifically seeks to evaluate different anomaly detection algorithms to determine the minimum number of password repetitions required for effective training, optimizing both system security and user convenience. Methodology. The study replicates and extends the evaluation procedure of Killourhy and Maxion, who provided a public dataset and a detailed protocol for analyzing keystroke dynamics. The algorithms are evaluated by varying the number of password repetitions used for training, with the aim of determining the optimal training size that balances security (lower EER) and efficiency (reduced user effort). Scientific Novelty. The scientific novelty of this research lies in its investigation of the trade-off between security and user convenience in keystroke dynamics-based authentication systems. While many studies have focused on improving the accuracy of anomaly detection, this research emphasizes the importance of minimizing the training burden on users by determining the minimum number of password repetitions required for stable performance. By focusing on training efficiency and computational resource optimization, this research advances the field of behavioral biometrics and contributes to the practical deployment of keystroke dynamics in real-world authentication systems. Conclusion. The study demonstrates that keystroke dynamics can significantly improve the security of twofactor authentication systems without imposing excessive burdens on users. The findings confirm that the Manhattan (scaled) and Outlier Count (z-score) algorithms perform relatively well, particularly when the training set size is small, which is critical for practical use in authentication systems where users may be unwilling to provide numerous password repetitions. This study not only replicates the results of prior research but also contributes new insights into optimizing the training process for keystroke dynamics-based anomaly detection. Future work may explore integrating keystroke dynamics with other biometric factors, such as mouse dynamics, to develop even more secure and user-friendly multimodal authentication systems. Furthermore, continuous authentication mechanisms represent an exciting direction for future research, providing ongoing verification of user identity throughout a session rather than solely at login.