Evaluating the Robustness and Transferable Adversarial Example Resistance of Multimodal Learning-based Intrusion Detection Systems Against Evasion Attacks

Intrusion Detection Systems (IDS) play a crucial role in safeguarding computer networks against malicious activities. However, current IDS models based on machine learning (ML) and deep learning (DL) encounter challenges in accurately classifying malicious network traffic, especially when confronted with adversarial perturbations and the transferability of adversarial attacks. To address these challenges, numerous studies have proposed and demonstrated the effectiveness of a novel multimodal approach, which leverages multiple sources of information to achieve better accuracy and the ability to identify more unknown attacks. In this study, we evaluate the robustness of multimodal learning-based IDS against transferable adversarial examples (AEs) generated by Generative Adversarial Networks (GANs). Additionally, we integrate adversarial training techniques to enhance the IDS’s capability to identify attack patterns with small perturbations. Our proposed strategy, MAT, achieves near-perfect detection rates across all attack types and the highest overall F1 score (0.7595), substantially outperforming all baselines under second-round adversarial attack conditions. While the attained results may not be as remarkable as desired, implementing the Multimodal approach makes a notable contribution to the field, paving the way for further research and advancements in addressing the challenges encountered by existing IDS models.