Adversarial Attack to Deceive One Stage Object Detection Algorithms

Abstract

In this paper, we are focusing to fool one stage object detection algorithms and propose a black box method to generate adversarial example such that it is imperceptible to human eyes and can fool one stage object detectors. We are generating random perturbation and scaling it on the basis of unsuccessful attack and maximum number of iterations. The generated perturbation is added to original image to generate perturbed image. After that the output of one-stage object detector is compared. We have defined three success scenarios, hiding objects, misclassification and count of objects, if attack achieved one of these scenarios, it will be considered as successful attack. The proposed work is evaluated on the basis of perceptibility, average number of iterations and convergence rate. The results show that we have achieved 98.05% convergence rate on 4.7 average number of iteration with PASS score of 1.94*10-2 on RetinaNet, 98.73% convergence rate on 4.68 average number of iteration with PASS score of 1.58*10-2 on Single Shot multi-box Detection (SSD) and 77.11% convergence rate on 6.08 average number of iteration with PASS score of 2.04*10-2 on You Look Only Once version 3 (YOLO V3) which shows the effectiveness of proposed attack.