Research On Generation Algorithm Of Adversarial Examples Based On Target Detection

Deep neural networks have been widely used in people’s lives.However,some criminals may use deep neural networks to identify collected user image data,thereby stealing personal privacy.In recent years,studies have found that deep neural networks are vulnerable to attacks from adversarial examples.By adding tiny perturbations that are imperceptible to the human eye,the classification results of deep neural networks are changed.Therefore,this thesis attempts to use adversarial example generation algorithms to protect image information and personal privacy.Based on the above situation,the research content of this paper is to use target detection or modify the target detection network,and propose an adversarial example generation algorithm.The main contents of this paper are as follows:(1)This thesis first analyzes the principle and process of the non-directional adversarial example generation algorithm DeepFool in dealing with two-class and multi-class problems,as well as the defects of the algorithm.Then,by improving DeepFool,a fixed-range adversarial example generation algorithm is proposed.(2)Aiming at the problem of using deep neural networks to steal image information,an image local information hiding algorithm is proposed.By combining the YOLO v3 target detection technology and the improved DeepFool,an image local information hiding algorithm is proposed.Compared with traditional methods of protecting local information in images,such as mosaic,blur,and partial occlusion,the algorithm in this paper is more concealed,while ensuring the usability of the image.(3)Two adversarial examples generation algorithms are proposed for Faster R-CNN.Faster R-CNN is a classic "two-stage" object detection algorithm.This paper analyzes the influence of the activation function in the deep neural network when the image produces perturbation,and the principle of the optimizer used for the update of the weight parameter in the neural network.And proposed two methods of using gradient and optimizer to modify the Faster R-CNN network model to construct the adversarial sample generation network for Faster R-CNN.Determine important parameters and optimizer types through experiments,and achieve the effect of recognizing multiple recognition objects of the object detection model as specified categories.