Abstract:

A camouflage method was developed for conducting physical adversarial attacks on vision-based vehicle detection systems to protect the privacy of personal vehicles. An objective function of the attack algorithm was optimized based on a 3D adversarial attack framework to enhance the effectiveness of 3D adversarial textures under multi-view and multi-scene settings. A weather-adaptive weighted hierarchical color mapping network was designed to enable adversarial textures to respond to weather parameters during the training process, further improving the robustness of physical-world attacks. Digital and physical experiments were conducted. The results show that the proposed algorithm reduces the average recall rate of detection by 49.4%. Therefore, the optimized adversarial textures demonstrate physical-world feasibility, achieving at least a 38.7% reduction in detection accuracy in real-world scenarios.

Key words: vehicle detection systems, physical adversarial texture attacks, deep learning, computer vision