부산대학교 도서관

The deep learning based object detection has been widely used. The adversarial attack algorithms reduce the accuracy of the recognition model by adding perturbation noise to the target image to achieve the purpose of hiding sensitive objects. Existing adversarial attack algorithms cannot balance the aggression and imperceptibility issues of adversarial perturbations. To address this problem, an Attention maps guided Adversarial Attack Method for object detection (AAAM) is proposed. This method first designs an initial perturbation region positioning strategy based on the Class Activation Mapping (CAM) algorithm and then uses the RPAttack’s gradient based contribution judgment perturbation region refinement strategy during the adversarial perturbation iteration generation phase, thereby achieving stronger attacks with a smaller perturbation region. The AAAM improves adversarial attack effectiveness through more efficient initial perturbation region selection. Compared to RPAttack, DPAttack and PGD, when attacking Faster RCNN AAAM and YOLOv4 AAAM achieves an improvement rate of over 7.2% and 16.2% in terms of the attack success rate (ASR) as an attack effectiveness metric.