Traffic adversarial example attack and defense method based on explainable artificial intelligence

An adversarial example attack method based on XAI was proposed for AI-based NIDS. By identifying critical perturbation features with XAI and applying targeted perturbations while preserving traffic functionality, malicious traffic was gradually altered to be misclassified as benign, achieving adversarial traffic sample attacks. This approach reduced the number of required perturbation features, enhancing attack stealthiness. The identified features showed consistency across classifiers, giving attack samples robust transferability. For defense, a defense method based on adversarial training was proposed to boost NIDS robustness. Experiments show high attack success and transfer rates, and the proposed defense method effectively lowers adversarial example attack success rates, enhancing system robustness.