Research on Local MI-PGD Adversarial Example Generation for YOLO Models

To enhance the adversarial robustness of YOLO-based object detectors, a localized MI-PGD approach is proposed for generating adversarial examples. The proposed method enhances the traditional PGD algorithm by incorporating a momentum-based iterative optimization strategy. Grad-CAM++ is then employed to identify the regions most sensitive to the model's detection. These regions are integrated with the MI-PGD algorithm to generate adversarial examples with localized perturbations, thereby improving both the effectiveness and stealth of the attack. Experimental results demonstrate that the proposed method effectively degrades the YOLO model's mean accuracy and recall rate under adversarial conditions. This study presents a novel approach for generating adversarial examples and offers valuable insights for improving the security and robustness of object detection algorithms.