Lichen Xia, Xing Gao, and Weisong Shi, University of Delaware
Object detection (OD) systems are fundamental to safety-critical applications such as autonomous driving and indoor mobile robots, yet remain vulnerable to latency-based attacks that degrade real-time performance. Existing attacks suffer from computational inefficiency and poor stealthiness, as phantom object generation disrupts original detections, making attacks easily identifiable. We propose Groundswell, a novel latency attack employing Regional Perturbation Balance (RPB) to inject adversarial phantom objects while preserving original detections. By leveraging spatial object distribution patterns, RPB generates masks that guide phantom placement into regions with minimal interference, enabling efficient optimization via Projected Gradient Descent. Evaluations across multiple OD architectures demonstrate that Groundswell reduces resource utilization by 50% and training time by 80% over existing methods, while successfully inducing latency delays that exceed real-time thresholds for safety-critical systems.
Open Access Media
USENIX is committed to Open Access to the research presented at our events. Papers and proceedings are freely available to everyone once the event begins. Any video, audio, and/or slides that are posted after the event are also free and open to everyone. Support USENIX and our commitment to Open Access.
