Enhancing Machine Learning-Based IDS for Vehicular Networks by Addressing Adversarial Attacks
Pegah Mansourian, Ning Zhang, Arunita Jaekel, Tim Allsopp · IEEE Transactions on Vehicular Technology · 2025
In Vehicular Ad Hoc Networks (VANET), Intrusion Detection Systems (IDS) are pivotal in ensuring secure communication among vehicles and infrastructure. These systems are tasked with identifying abnormal behavior, including malicious attacks or unauthorized access, within the dynamic VANET environment. However, the advent of Adversarial Examples (AE)-inputs crafted to deceive machine learning models-has introduced new challenges to IDS effectiveness. To address this, researchers are exploring methods to integrate adversarial examples into IDS frameworks to enhance security and resilience against attacks. This paper presents a comprehensive framework designed to bolster the robustness of IDS within VANETs against adversarial attacks. Leveraging adversarial examples and employing a rigorous verification process, our methodology systematically fortifies classification models against potential threats. Through iterative data generation, verification, and model adjustment stages, we ensure the creation of verified adversarial examples with an optimal level of added perturbation, culminating in a final robust model, calledAdversarially-Fortified Vehicular IDS (AFV_IDS), capable of confidently discerning between normal and attack messages. The integration of adversarial training techniques further enhances the model's resilience, addressing previously unseen blind spots and adjusting decision boundaries to accommodate adversarial instances. Our framework offers a holistic solution to enhance the security of vehicular networks, thereby contributing to safer and more reliable transportation systems.