A Systematic Study of Adversarial Attacks Against Network Intrusion Detection Systems

Network Intrusion Detection Systems (NIDSs) are vital for safeguarding Internet of Things (IoT) networks from malicious attacks. Modern NIDSs utilize Machine Learning (ML) techniques to combat evolving threats. This study systematically examined adversarial attacks originating from the image domain against ML-based NIDSs, while incorporating a diverse selection of ML models. Specifically, we evaluated both white-box and black-box attacks on nine commonly used ML-based NIDS models. We analyzed the Projected Gradient Descent (PGD) attack, which uses gradient descent on input features, transfer attacks, the score-based Zeroth-Order Optimization (ZOO) attack, and two decision-based attacks: Boundary and HopSkipJump. Using the NSL-KDD dataset, we assessed the accuracy of the ML models under attack and the success rate of the adversarial attacks. Our findings revealed that the black-box decision-based attacks were highly effective against most of the ML models, achieving an attack success rate exceeding 86% across eight models. Additionally, while the Logistic Regression and Multilayer Perceptron models were highly susceptible to all the attacks studied, the instance-based ML models, such as KNN and Label Spreading, exhibited resistance to these attacks. These insights will contribute to the development of more robust NIDSs against adversarial attacks in IoT environments.