Research On The Application Of Generative Adversarial Network In Adversarial Attack Defense And Super-resolution

In the current upsurge of artificial intelligence,the proposal of generative adversarial network GAN meets the research and market application needs of many disciplines,and instill new development impetus for these researches and applications at the same time.In view of the advantages of GAN in data expression and distribution learning,through in-depth research and analysis of the current development of confrontation attacks and super-resolution,this paper applies GAN to the field of adversarial attack defense and super-resolution.In the aspect of counterattack defense,this paper combines the generative adversarial network with the existing attack algorithm to make the network have a deeper understanding of the image distribution,and propose an adversarial attack defense model AC-Def GAN.In this paper,we use the adversarial attack model to generate attack samples as training samples,and the classification of samples generated by the generator is used to guide the training process of GAN,and then we add the conditional constraints to stabilize the model training process.Once our model is trained,the classifier of this model can effectively defend the corresponding adversarial attacks,and its training process does not need to know the structure and parameters of the target model.The experimental results on the MNIST,CIFAR-10 and Image Net datasets prove the effectiveness of AC-Def GAN.In terms of super-resolution reconstruction,the super-resolution generative adversarial network SRGAN has a better reconstruction effect.However,the reconstructed image is still not realistic enough and there is still a gap between the reconstructed image and the real image.In this paper,we improve and optimize the SRGAN algorithm thoroughly study three key components of SRGAN.We improve the loss function and network structure,and propose an improved super-resolution reconstruction model SRRWGAN based on generative confrontation network.Compared with SRGAN,this model removes all the BN layers in the generator network to improve generalization ability and reduce computational complexity,and RRDB is introduced to improve generator network to improve network performance.In addition,by combining with Wasserstein GAN,we optimize the objective function of the generated network to improve the training stability of the network,and the discriminant network was improved correspondingly improved.The experimental results on data sets such as DIV2K?Set5?Set14?BSD100 and Urban100 verify the reconstruction effect of the model proposed in this paper.