Investigating the Robustness Against Transferable Adversarial Attacks of Learning-Based Network Intrusion Detection System
Phuoc-Sang Dang, Phan The Duy, Van-Hau Pham · 2024
Intrusion Detection Systems (IDS) play a crucial role in safeguarding networks against increasingly sophisticated cyber threats. Adversarial attacks involve intentionally crafting malicious inputs to exploit vulnerabilities in learning-based IDS models, causing them to misclassify or overlook suspicious activity. Adversarial transferability refers to the ability of these malicious inputs to deceive not just the model they were designed for but also other models with similar architectures or characteristics. This study presents a novel methodology for evaluating the effectiveness and robustness of IDS models by examining their vulnerability to various evasion attacks and the transferability of adversarial examples across models with different levels of complexity. We compare machine learning (ML) and deep learning (DL) models, each represented in both lower and higher complexity versions. The methodology includes well-established evasion attack techniques such as Fast Gradient Sign Method (FGSM), Projected Gradient Descent (PGD), and Momentum Iterative Method (MIM), tested in both targeted and untargeted modes. Our experiments, conducted on a comprehensive and realistic network traffic dataset, demonstrate that while DL models typically achieve strong performance, they are more susceptible to adversarial attacks compared to traditional ML models. Furthermore, the results indicate that adversarial examples generated by lower-complexity models are often more effective at misleading higher-complexity models in black-box scenarios. This research offers critical insights into strengthening the robustness of learning-based IDS models against adversarial threats and provides a valuable foundation for future IDS advancements.