MARAGAN-IDS: A Bidirectional Residual GAN for Imbalanced Intrusion Detection

In the context of increasingly frequent and complex cyberattacks, network intrusion detection systems (NIDS) face significant challenges due to class imbalance in data. Traditional Wasserstein Generative Adversarial Networks (WGANs) encounter two major issues in intrusion detection: difficulty in generating semantically rich minority class samples and a lack of diversity in generated outputs, often leading to mode collapse. To address these problems, this paper proposes an improved intrusion detection framework, MARAGAN-IDS, which centers on a bidirectional residual mechanism in the generator to enhance feature representation and improve the semantic consistency and quality of minority class samples. Moreover, to further mitigate the homogenization of generated data, we introduce cosine similarity into the WGAN loss function to construct a hybrid loss, thereby enhancing both the diversity and realism of the generated samples. Extensive experiments conducted on two widely used intrusion detection datasets, NSL-KDD and CIC-IDS2017, demonstrate that compared with state-of-the-art models, MARAGAN-IDS achieves superior performance in terms of detection rate, precision, and F1-score.