AE-Multi-WGAN: A Robust Multi-Attack Detection IDS for Imbalanced Datasets

Intrusion Detection Systems (IDS) are vital for identifying malicious activities in complex network environments. Numerous approaches, from traditional Machine Learning (ML), statistical techniques, and Deep Learning (DL) models, have been developed to detect attacks using IDSs. However, existing approaches suffer from high false positive (FP) rates, lower accuracy, and heavy computational load. Amongst them, Generative Adversarial Networks (GANs) have gained significant interest due to their ability to model complex data distributions and generate realistic synthetic samples. However, GAN-based IDS methods still face key challenges, such as class imbalance, mode collapse, training instability, and difficulty in capturing diverse attack patterns, which limits their generalization and reliability in practice. To address these issues, we propose an Ensembled AutoEncoder Multi-Wasserstein GAN (AE-Multi-WGAN) scheme, integrating multiple generators and discriminators with an AutoEncoder (AE) in an ensemble DL framework to improve attack detection. The AE enables denoising and feature compression, while the WGAN component ensures stable training and high-quality synthetic sample generation. Validated on the NSL-KDD and UNSW-NB15 datasets, the proposed AE-Multi-WGAN demonstrates superior precision, recall, and F1-score compared to baseline methods, effectively enhancing IDS performance, reducing FP, and improving generalization to unseen attacks.