Exploring Autoencoders for Anomaly Detection in Intrusion Detection Systems

This paper presents the development and evaluation of an Intrusion Detection System (IDS) based on a Denoising Autoencoder (DAE). Three benchmark datasets—NSL-KDD, UNSW-NB15, and CTU-IoT—were used to assess the model’s capability to detect anomalous network behavior across diverse environments. The proposed DAE achieved 86.1% accuracy and an 87.1% F1-Score on NSL-KDD, 94.8% accuracy on UNSW-NB15, and up to a 78.6% F1-Score on CTU-IoT depending on the anomaly threshold. These results are comparable to or exceed several autoencoder-based IDS approaches in the literature. Overall, the findings highlight the importance of threshold selection, dataset characteristics, and representation quality in anomaly-based intrusion detection.