Abstract

Present-day cyberattacks, including phishing, brute-force attacks, and credential stuffing, highlight the limitations of traditional authentication techniques such as passwords and PINs, which are susceptible to weak user credentials and password reuse. This study proposes a continuous, non-intrusive authentication method based on keystroke dynamics, a behavioral biometric that analyzes precise typing patterns. We introduce a deep learning model leveraging a multi-input long short-term memory (LSTM) architecture augmented with an attention mechanism to capture complex temporal typing behaviors. Evaluated on the KeyRecs dataset (2023), our model achieves an accuracy of 99.34%, a false acceptance rate (FAR) of 0.01%, a false rejection rate (FRR) of 0.66%, and an equal error rate (EER) of 0.01%. These results significantly outperform the Carnegie Mellon University (CMU) baseline (93.2% accuracy, 3.9% EER) and comparable models (EERs of 0.75–10%). The proposed system demonstrates keystroke dynamics as a robust alternative for standardizing authentication strategies, with future work targeting enhanced robustness through multi-modal biometrics and equitable performance across diverse demographics using fairness metrics.