Investigating Stealthy Latency Attacks Against Camera-Based Object...
summary
Groundswell crafts a single universal sticker/projection perturbation using Regional Perturbation Balance and PGD to inject phantom proposals into camera frames while preserving original detections; evaluation on Berkeley DeepDrive with YOLOv5/8/11 shows a ~180× jump in NMS candidates, NMS latency increases ~117%, recall drops to 18% at aggressive settings but stays ~38% at T=0.4 and preserves key detections at T=0.8; optimization cuts CPU use by 50% and training time by 80%.
AI-written summary · about
related talks
11mWIP: Evaluating the End-to-End Impact of False Localization Attacks on vSLAM-...
55mPerception Deception: Physical Adversarial Attack Challenges
13mBeyond the Glow: Understanding Luminescent Marker Behavior Against Autonomous...
40mSecuring Tesla and Mobileye from Split-Second Phantom Attacks
09mWIP: Hijacking Attacks on UAV Follow-Me Systems in Realistic Scenarios
10mLightning Talk: Ball Tracking and Detection in Soccer Videos - Comparison of...
21mSoK: Efficiency Robustness of Dynamic Deep Learning Systems
01h 06mSelf-Driving and Connected Cars: Fooling Sensors and Tracking Drivers