When Collaboration Betrays: The Evolving Attack Surface of...
summary
Claims that a malicious collaborator can craft shared feature maps to produce 0.5 m post-perturbation of an object’s reported position and, in CARLA scenarios, trigger hard braking or collisions by cascading through motion prediction. Demonstrates a practical attack and evaluates a usable defense: localized anomaly detection around critical objects increases sensitivity versus whole-map checks. Identifies trust poisoning where a hollow or rear-mounted adversarial object makes one vehicle’s view appear anomalous so its contributions get removed; proposes self-reporting uncertainty and shows real-world prototype experiments.
AI-written summary · about
related talks
13mWIP: Learning Adversarial Attacks on Adaptive Traffic Signal Control Systems...
28mSecurity of Machine Learning
34mHacking the Blackbox
35mHacking AI Systems: How to (Still) Trick Artificial Intelligence
21mAdversarial Attacks
55mPerception Deception: Physical Adversarial Attack Challenges
11mWIP: Evaluating the End-to-End Impact of False Localization Attacks on vSLAM-...
46mA Short Introduction to Hacking Artificial Intelligence