Adversarial Patches for Fooling Positive–Negative Obstacle Detectors on Unstructured Roads

Yijian Duan, Liwen Meng, Jihong Zhu, Yanmei Meng, Yuan Liang, Jinlai Zhang, Jiachun Wu

Abstract: Adversarial patches represent a critical form of physical adversarial attacks, posing significant risks to the security of neural network-based object detection systems. Previous research on adversarial patches has predominantly focused on pedestrian detection, facial recognition, and vehicle detection, with limited attention to the detection of positive and negative obstacles on unstructured roads. Moreover, prior studies typically optimize perturbation information while fixing parameters such as the patch’s position and rotation angle or manipulate parameters such as position and rotation angle while fixing the perturbation information to generate adversarial patches. In this context, we propose a position, rotation angle, and perturbation pixel values multiadversarial patch attack method based on a simulated annealing improved differential evolution (SADE-PRP) algorithm, designed to deceive unstructured positive and negative obstacle detection systems. First, we propose a position, rotation angle, and perturbation pixel values (PRP)-guided multiadversarial patch framework for visible light-based positive–negative obstacle detection systems. This framework simultaneously considers three features of the adversarial patch: position, rotation angle, and perturbation pixels, as opposed to manually setting these parameters like most previous works. Second, we employ a more robust-based simulated annealing improved differential evolution (SADE) algorithm, which effectively improved the robustness, achieving higher attack success rates (ASRs) in black-box attack scenarios. Finally, we constructed a dataset of positive–negative obstacles on unstructured roads. Then, we conducted extensive experiments in both digital and physical environments to demonstrate the superiority of the proposed method.

External IDs:doi:10.1109/jsen.2025.3586858