Implementation of Deep Learning Based on Long Short-Term Memory (LSTM) for Continuous Authentication Through Keystroke Dynamics Analysis : A Simulation Study of Students at WR Supratman 1 Private Senior High School, Medan

Abstract

Conventional password-based authentication verifies a user's identity only once, at the point of entry, which leaves the session open to hijacking after a successful login. Keystroke dynamics offers a continuous alternative built on typing behavior that is distinctive to each person, but its accuracy depends on how well a model captures the temporal dependencies between keystrokes. This study implements a Long Short-Term Memory (LSTM) deep learning model for continuous authentication based on keystroke dynamics, compares it with conventional machine learning models, and explores how users perceive such a system. We used a mixed-methods design with realistic simulated data representing 42 eleventh-grade students at WR Supratman 1 Private Senior High School in Medan. Each participant produced dwell time and flight time samples from a fixed passphrase and from free text, and 10 participants took part in semi-structured interviews. The LSTM model reached the highest classification accuracy (96.8%), ahead of Random Forest (91.5%), Support Vector Machine (89.2%), and k-Nearest Neighbor (85.6%). Its Equal Error Rate (EER) was 2.75%, compared with 6.80–10.85% for the baselines. Most interviewees felt safer under continuous authentication, although many were concerned about privacy because of the constant monitoring. We conclude that LSTM captures the temporal patterns of keystroke dynamics markedly better than the baseline models, but its deployment should be accompanied by transparency toward users to maintain trust and acceptance.