Idea Description:

Summary: Collaborative vehicular perception enhances autonomous driving by enabling vehicles to share sensory data, improving situational awareness. However, this shared perception framework is susceptible to adversarial attacks that can fabricate or perturb data, compromising the safety and reliability of autonomous systems. Existing defense algorithms either have high latency or low generalizability to multiple attack methods. We would like to introduce uncertainty awareness into the defense mechanism.

Intuition: Imagine someone tells you there’s an obstacle ahead while you’re driving. If you clearly see nothing in front of you, you’ll likely reject their suggestion because you’re confident in your own observation. However, if you are nearsighted or looking elsewhere, you might accept their suggestion since you’re uncertain about your perception. This uncertainty-aware analysis can serve as a guide for deciding whether to accept or reject information from other agents.