Research On Generation Model Of Perceptually Similar Image Classification Adversarial Example

In the field of image classification,adversarial examples refer to the images that slightly modified from the original images in human perception,which could cause the classifiers to generate incorrect classification results.Most of the existing adversarial example generation algorithms in image classification use the iterative method to modify the original image multiple times based on information such as gradients,until an adversarial example is constructed,which requires multiple backward and forward calculations,resulting in excessive time consuming of adversarial example generation.The generator-based adversarial example generation model requires only one-time forward calculation to generate adversarial example once the model has been trained,which takes a short time.However,the example generated by the existing generator-based models are clearly different from the original image in human perception.In order to reduce the generation time of the adversarial example,and reduce the difference between the adversarial example and the original image in human perception,this thesis proposes novel generator-based adversarial example generation models:GPAE and t-GPAE,for non-targeted and targeted adversarial attacks respectively.Compared with other generator-based adversarial examples generation models,they could effectively improve the similarity between adversarial example and the original image,while maintaining the fooling ratio.The GPAE model and the t-GPAE model no longer consider the generation of adversarial examples as a perturbation superposition of the original image,but as an image enhancement operation on the original image so that it could make the changes applied to the original image distributed in areas that human can hardly noticed.At the same time,GAN and the perceptual loss function are introduced into the loss function,and improvements are made to increase the reality of the generated adversarial example and the similarity with the original image in human perception.This thesis also uses a multi-classifier loss function to train the generator to generate adversarial examples that could attack multiple classifiers to improve the efficiency of adversarial attacks.For t-GPAE model,which based on GPAE,we also improved the model structure and introduced WGAN-GP to improve model stability and speed up model convergence.In this thesis,experiments are performed on GPAE and t-GPAE models on standard datasets.Experimental results show that compared with other generator-based adversarial example generation models,the SSIM indicators have been improved.At the same time,this thesis also carried out experiments on improvements we made in the GPAE model and the t-GPAE model,and experimental results proved the effectiveness of each improvements we made.This verifies that GPAE and t-GPAE model could generate more similar adversarial examples to the original images in human perception while keeping the fooling ratio in non-targeted and targeted attacks.