Enhancing the Performance of Noise-Resistant Autoencoder-Based Network Intrusion Detection Systems

Abstract

With the growing complexity of real-world cyber threats, organizations face increasing challenges in protecting their networks from attacks that often blend in with normal traffic. In practical settings, network data is frequently noisy and only a small portion of it is labeled, making intrusion detection a difficult task. Semi-supervised deep learning models, particularly autoencoders, have gained popularity for detecting unusual activity in network traffic. However, standard autoencoders often struggle to adapt to the unpredictable and noisy conditions found in real network environments, giving rise to overfitting and a higher rate of false alarms. In this work, we present a noise-resisting autoencoder, which is the result of a rigorous experiment on different noise types, including less conventional ones, noise levels, and other key hyper parameters. In this study, we used ANOVA and HSD statistical methods, which offered a solid and interpretable basis for a multi-layered, evidence-based understanding of how each hyper parameter shapes the model’s behavior, while providing practical guidance for building more reliable and noise-resistant autoencoder-based intrusion detection systems. Our model was tested on the NSL-KDD achieving an accuracy of 91.64% outperforming state-of-the-art semi-supervised autoencoders models.