A Low-Resource Convolutional Autoencoder Approach for Anomaly Detection in Web-Based Applications

As web applications become essential business tools, website security is critical due to attacker targeting of valuable information. Anomaly detection, a crucial daily task, mitigates system damage and its potentially dangerous consequences. Numerous machine learning studies explore this area, with deep learning playing a significant role. Motivated by the need for low-resource deep learning models suitable for edge computing, this study develops three well-known deep models (Vanilla Autoencoder, Deep Autoencoder, and Convolutional Autoencoder) with a focus on minimizing the number of parameters. A real-world dataset of HTTP/HTTPS requests from a university portal (with complex payloads due to diverse features and operations) is used for evaluation. For comparison, a simpler in-lab dataset is also included. Our findings confirm the suitability of deep learning for anomaly detection and identify CAE as the most effective technique in this scenario. CAE requires the fewest training parameters and achieves reasonable training time without GPUs. A key contribution of this work is the application of these research results to a real system using the real dataset for efficient anomalous request detection.