GenAtt NIDS Attack: A Practical Generative Adversarial Attack Against Network Intrusion Detection Systems

Network Intrusion Detection Systems (NIDS) have made tremendous progress by combining Deep Learning (DL) and Machine Learning (ML) techniques, and they currently perform well against both known and unexpected threats. However, because of intrinsic flaws, adversarial attacks may reduce their detection effectiveness. In this study, an attention-based generative method for developing adversarial assaults utilizing Generative Adversarial Networks (GANs) is presented. Our approach avoids IDS detection by leveraging attention mechanisms' sequential memory capabilities and producing semantically meaningful modified packets. We address the heterogeneous nature of attack packets with normalized evaluation, enabling meaningful inferences for cybersecurity experts. Using a black-box adversarial attack with Active Learning and request-and-answer queries, we achieved a practical analysis. Compared to prior work showing a 15.3 % drop in IDS detection rates, our method resulted in an 11.3% drop-a slight but considerable reduction-while using only 20 % of the queries. This highlights the efficiency of our approach in reducing anomalous query volumes and maintaining realistic network presence.