Adaptive Continuous Adversarial Training (ACAT) to Enhance ML-NIDS Robustness

In this paper, we extend our previously proposed Adaptive Continuous Adversarial Training (ACAT) method beyond the problem-space of SPAM filters to encompass the featurespace of Machine Learning (ML)-based Network Intrusion Detection Systems (NIDS). ACAT continuously improves model robustness by incorporating adversarial samples detected from the ML-NIDS input network traffic into the training process. Problem-space evasion attacks rely on inverse-feature mapping, where modifications to real-world objects result in targeted perturbations within the feature vector after feature extraction. Therefore, protecting ML systems against feature-space attacks inherently provides defense against both feature-space and problem-space evasion attacks. Our results demonstrate that ACAT significantly reduces adversarial sample detection time compared to traditional techniques. Moreover, the performance of the ML-based NIDS under attack improved dramatically, with accuracy increasing from 50.92 % to over 99 % after only three retraining sessions.