Adversarial attacks against network intrusion detection systems: Bridging the gap between theoretical vulnerabilities and practical constraints

Network Intrusion Detection Systems (NIDS) have become critical components of cybersecurity infrastructure, with machine learning-based approaches achieving detection accuracies exceeding 99% on benchmark datasets. However, the emergence of adversarial machine learning has exposed fundamental vulnerabilities in these systems, where carefully crafted perturbations can evade detection with success rates surpassing 97% in laboratory conditions. This comprehensive literature review examines adversarial attacks against NIDS. Our investigation reveals a critical disconnect between theoretical attack capabilities and practical implementation constraints. Examining attack methodologies (gradient-based, GAN-based, black-box, and adaptive), the feature-space to network-packet translation challenge, defense mechanisms, and dataset limitations, this review applies a constraint-aware framework to quantify practical attack viability. Protocol compliance, semantic preservation, and infrastructure filtering render up to 80.3% of adversarial examples invalid under real-world constraints. Defense mechanisms including adversarial training, ensemble methods, and moving target defense face scalability barriers limiting production deployment. These findings suggest that many published attacks substantially overstate real-world vulnerabilities, providing a more nuanced understanding of practical threat landscapes for ML-based network security systems.