Skip to content
LowResearchPeer-reviewed

Generative Red Teaming for Robust Network Intrusion Detection: A Constrained Adversarial Approach

Published
Record updated
View JSON

Summary

Researchers introduce PerturbLite, a generative red teaming framework for assessing ML/DL-based network intrusion detection systems (NIDS) under realistic constraints. Its autoencoder-based model produces minimally perturbed, functionality-preserving adversarial packets that evade detection, often altering less than 1% of the original content. The authors report superior evasion performance over existing adversarial methods on real-world intrusion datasets and say the framework exposes systemic NIDS vulnerabilities.