Vulnerability Assessment of Adversarial Image Attacks in a Miniaturized Autonomous Vehicle Testbed without Relying on Way-Point Information

This paper presents a real-time perception and control framework to assess adversarial image attacks on an autonomous vehicle testbed. The system integrates 2D LiDAR with a YOLOv5n-based detector in a multithreaded architecture to execute a staged multi-stop navigation routine with turns and visual alignment before each stop. Baseline experiments show reliable operation without requiring way-point information, with consistent stopping accuracy and path completion across repeated runs. Instead of using traditional digital overlays, we evaluated our testbed on adversarial patterns on a physical traffic sign. To assess vulnerability against adversarial images, specific adversarial variants of traffic signs were tested. This cause marked reductions in detection range and even missed recognitions. To restore performance without changing sensing, fusion, or control, we apply adversarial learning to fine-tune YOLOv5n. The retrained model recovers detection reliability across all variants and returns the vehicle to normal closed-loop behavior without relying on waypoint information.