ADTA: Attention Disruption Transferable Attack Against Object Detection

In recent years, it has been demonstrated that object detection models exhibit significant vulnerabilities when confronted with carefully crafted adversarial attacks. However, most existing adversarial attack methods targeting object detectors are primarily focused on white-box settings, where the generated adversarial samples generally suffer from poor transferability across different models. To solve this problem, we propose an attack method named ADTA (Attention Disruption Transferable Attack) from the perspective of disrupting the attention of object detection models. ADTA first employs EigenCam to obtain the class activation maps of the object detection model for the original input images. Subsequently, it swaps the pixels with high and low attention weights in the original input images. These modified images are then fused with the original images in a master-slave manner at a certain ratio to construct an augmented dataset. Finally, adversarial samples are generated through multiple iterations. Empirical evaluations on COCO dataset demonstrate that ADTA could achieve significantly better transferability than existing attack methods.