CIAK-CP: Camera feed Injection AttacK in Collaborative Perception

Collaborative Perception (CP) enhances the capabilities of singlevehicle perception by producing collaborative outputs derived from information shared among vehicles. This helps overcome critical issues such as occlusions and sensor range limitations. However, as Connected Autonomous Vehicles (CAVs) achieve higher levels of autonomy, competition for resources becomes a realistic concern, and malicious CAVs may fabricate false information to gain an advantage despite being trusted nodes in the collaborative network. This paper proposes CIAK-CP, a framework for data-injection attacks against intermediate, camera-based collaborative perception systems. Our framework enables the extraction of vehicle image snippets from normal driving scenarios, which can later be injected into camera feeds and recognized as authentic vehicles by the perception module, even when they are not present in the real world. We present a detailed analysis of two safety-critical scenarios, a comprehensive evaluation of the attack’s effectiveness on 205 samples, and a discussion of state-of-the-art defenses and future directions.Our results reveal a severe limitation in the current CP defense research, and point out a possible direction for researchers to focus on.