Adversarial examples with transferred camouflage style for object detection. Issue 1 (January 2021)

• Record Type: Journal Article
• Authors: Deng, Xiaotong; Fang, Zheng; Zheng, Yunfei; Wang, Yang; Huang, Junming; Wang, Xiaobing; Cao, Tieyong
• Is Part Of: Journal of physics. Volume 1738:Issue 1(2021)
• Publication Date: 2021-01
• Subjects: Physics -- Congresses; 530.5
• DOI: 10.1088/1742-6596/1738/1/012130
• Languages: English
• ISSNs: 1742-6588

Abstract

Abstract: Most of existing adversarial examples attacking methods for object detection models aim at generating subtle perturbation which is invisible to human vision. However, some perturbations with stronger intensity are equally effective and can even make the objects in adversarial examples more invisible. In this paper, an attack model based on the Generative Adversarial Network and style transfer method is designed to craft and add perturbation with camouflage style to the object area in the image. Experiments on PASCAL VOC 2012 dataset demonstrated that Adversarial examples generated from our model can attack both proposal-based detectors and regression-based object detection models effectively, and reduce the visual saliency of the objects.