SIGNATURE WORK
CONFERENCE & EXHIBITION 2024

As autonomous vehicles gain popularity, ensuring their safety and security becomes important. Multi-agent Collaborative Perception (MACP) enhances autonomous cars’ perception capabilities, contributing to their safety. Despite its benefits, MACP is vulnerable to adversarial attacks, wherein malicious agents introduce adversarial examples to compromise the perception of recipient vehicles. This paper investigates the vulnerability of MACP to adversarial attacks, analyzing both white-box (where the attacker has access to the target model’s gradients) and black-box scenarios (where the attacker lacks such access). We introduce an Agent-Aware Defense System designed to mitigate these vulnerabilities in generic scenario. Through the application of the Fast Gradient Signed Method and Transferability techniques, we generate perturbations to assess the robustness of MACP models under attack. Our results indicate that white-box attacks are more effective than black-box attacks and that early-stage collaborative methods are particularly susceptible to adversarial manipulation. This study shows the necessity of integrating robust defense mechanisms in the development of collaborative perception frameworks for autonomous vehicles.