GDSA : Research on Goal‐Directed Single‐Step Noise Adversarial Attack in the Field of Autonomous Driving Safety

The use of Vehicle Inspection technology in recognizing images of road vehicles has increased the risk of adversarial attacks, resulting in reduced model accuracy. While many researchers have attempted to train the model to defend itself, this often leads to degraded detection performance. In this paper, we propose a new approach called Goal‐Directed Single‐Step Attack (GDSA), which aims to improve the model's stability and anti‐interference capability. Based on the large‐scale autonomous driving BDD100K dataset, GDSA designs transferable adversarial attack algorithms to attack the original intelligent driving scenario, and proposes the concept of attack algorithm generalization, combining the actual scenario to analyze the importance of achieving the attack goal on the basis of reduced perturbation. Our approach achieves an overall average mAP decrease of 15.85 across different target models and an mAP value of 46.32 after defense training on the validation set, demonstrating a qualitative improvement in model stability compared to other attack methods. The experiments also further conduct on the PASCAL VOC 10% dataset, achieving an mAP value of 49.48 after defense training on the validation set, which highlights the excellent performance of GDSA. GDSA not only provides strong resistance against adversarial attacks, but also offers potential for future expansion and application to other scenarios.