A new attack can slow a self-driving car’s camera perception without making the tampering obvious. Researchers at the University of Delaware developed Groundswell, a latency attack that injects adversarial phantom objects into camera-based object detection while preserving the original detections.
Latency attacks degrade the real-time performance that safety-critical systems rely on, but earlier versions were easy to spot. Phantom objects generated by prior techniques disrupted the legitimate detections around them, giving the game away. Groundswell uses a method the team calls Regional Perturbation Balance to place phantoms in parts of the image where they interfere least with genuine objects, guided by spatial object distribution patterns and optimized with projected gradient descent.
The result is a stealthier attack that still forces the perception pipeline to spend compute on fake inputs, delaying detection of real pedestrians, vehicles, and obstacles. Evaluations across multiple object detection architectures show Groundswell degrades real-time performance more efficiently than earlier latency attacks while remaining hard to identify.
For automakers building Level 2-4 systems, the work is a reminder that sensor security includes timing, not just spoofed pixels. The paper was presented at VehicleSec 2026.
