Deep Reinforcement Learning-Based Adversarial Defense in Vehicular Communication Systems

One of the key concerns related to the pervasive integration of artificial intelligence and machine learning (AI/ML) models in vehicular-to-everything (V2X) communication systems pertains to adversarial attacks, which may lead trained models to exhibit undesirable behaviors. As security and user safety are tightly coupled in V2X, ensuring the resilience of AI/ML models against adversaries becomes indispensable. However, addressing adversarial attacks poses a challenging task, requiring appropriate countermeasures to elevate the trustworthiness of the targeted AI/ML models. In this paper, we propose a deep reinforcement learning (DRL)-based approach to defend against two data poisoning attacks, namely label-flipping and policy induction. Extensive evaluation with the aid of an open-source dataset demonstrates that our scheme outperforms benchmark classifiers, achieving significantly superior detection performance in the presence of label-flipping attacks. The effectiveness of our DRL-based approach is also showcased under different adversarial strategies in the policy induction attack.