Adaptive Threshold-Based Autoencoder Model for Enhancing Security through Anomaly Detection in E-Learning Environments

Abstract

Securing the e-learning platforms has become increasingly difficult as user activities grow more complex and cyber threats become more advanced. Conventional detection techniques often fail to capture subtle irregularities in behavior, which can result in a higher number of false alarms and reduced accuracy. To overcome these limitations, this study adopts an autoencoder model enhanced with an adaptive threshold mechanism. The model first learns the pattern of normal user activity and then identifies unusual behavior by measuring reconstruction errors—where typical patterns are reconstructed accurately, while anomalies produce larger deviations. Unlike fixed thresholds, the adaptive approach adjusts dynamically based on the distribution of these errors, allowing for more precise detection. The experimental findings show strong performance, achieving 99.35% accuracy, 98.84% precision, 100% recall, and an F1-score of 99.41%, clearly surpassing traditional methods. Overall, the results suggest that this approach is highly effective for detecting anomalies and can significantly strengthen security in modern, data-driven elearning environments.