Reproduce “CP-FREEZER: Latency Attacks against Vehicular Cooperative Perception” on OpenResearch

About this paper

Chenyi Wang, Ruoyu Song, Raymond Muller, Jean-Philippe Monteuuis, Z. Berkay Celik, Jonathan Petit, Ryan Gerdes, Ming Li

Researchers at the University of Arizona, Purdue University, Qualcomm, and Virginia Tech introduced CP-FREEZER, the first latency attack against vehicular Cooperative Perception (CP) systems. This attack effectively achieved over a 90x increase in end-to-end perception latency, pushing per-frame processing times beyond 3 seconds, and consistently inflated pre-NMS bounding box proposals by an average of 46.19x.

Key insights

• The quadratic time complexity of the NMS stage in intermediate fusion CP systems represents a critical, previously unexploited bottleneck highly susceptible to latency attacks.
• Current integrity-focused defenses are ineffective against latency attacks and can, in fact, amplify the induced delays, highlighting a fundamental gap in existing autonomous vehicle security paradigms.
• Effective latency attacks against CP can be achieved by crafting adversarial perturbations directly at the BEV feature level and synchronizing them despite asynchronous communication using feature warping.

Summary by alphaXiv.