Adversarial Red Teaming for NIDS:Model-Agnostic Physical-Space Attacks

Abstract Adversarial examples have been widely studied across domains such as image recognition and speech processing, but their implications for Network Intrusion Detection Systems (NIDS) remain underexplored. This paper introduces the NIDS Robustness Toolbox (NRT), a model-agnostic tool designed to generate adversarial samples for machine learning-based NIDS by directly manipulating network traffic in the physical space. Unlike traditional approaches, our method requires no information about the processing pipeline or underlying model architecture, making it adaptable and highly practical for real-world red teaming. The NRT facilitates the creation of adversarial traffic that disguises malicious intent within realistic network behavior, targeting both flow-based and payloadbased NIDS models. Experimental evaluations demonstrate that these adversarial samples lead to significant declines in detection performance, assessed across multiple metrics including precision, recall, F1-score, evasion, and misclassification rates.These findings reveal substantial vulnerabilities in current NIDS configurations, underscoring the critical need for more resilient, adaptable defenses to withstand sophisticated adversarial threats in real-world deployments.