Image Processing Algorithm Based On Adversarial Deep Learning

As an intuitive and efficient way of expression,images play a very important role in people's production and daily life.Therefore,image processing technology has also received widespread attention in the academic community.In recent years,research on adversarial deep learning has developed rapidly,injecting new vitality into image processing.GAN(Generative Adversarial Network)is the core of adversarial deep learning,which with excellent learning ability and fitting ability,are especially good at processing image-related tasks.This thesis focuses on the application of GAN in image-to-image translation tasks and adversarial attack tasks against image recognition networks.Due to different image processing tasks,the specific architecture of the generation adversarial network and the application mechanism introduced also have their own focuses.This thesis proposes two image processing algorithms based on adversarial deep learning,namely MMA-CycleGAN(Multi-head Mutual-attention Cycle Consistence GAN)and WP-AdvGAN(Adversarial Attack GAN with Weighted Perturbations).Details are as follows:Image-to-image translation is the transformation of images from one expression(source domain)to another expression(target domain).It has broad application prospects in industrial design,artistic creation and so on.The existing algorithm for unsupervised image-to-image translation,CycleGAN only uses dual learning to constrain the fitting process of the network.As a result,the generator cannot quickly and accurately learn key features in the target domain,so the quality of the translated images need to be further improved.Based on this,this thesis introduces the Mutual Attention mechanism into CycleGAN for the first time,and proposes an improved algorithm MMA-CycleGAN for unsupervised image-to-image translation tasks.The Mutual Attention mechanism is applied to establish a long-distance relationship between the source domain and the target domain,and to use attention-driven modeling to enable images of the source domain to directly learn key features in the target domain.For large-size images,the basic Mutual Attention mechanism is further improved to a Multi-head Mutual Attention mechanism to save more computer memory costs.Experimental results show that compared with other unsupervised image-to-image translation algorithms,MMA-CycleGAN can more efficiently establish the mapping of key features between two image domains,and significantly improve the quality of translated images.Image recognition technology has been widely used in face recognition and autonomous driving.But research shows that adversarial examples with malicious perturbations are able to attack image recognition networks and make them misjudge.At present,most of the attack strategies represented by AdvGAN construct global perturbations.In spite of a higher attack success rate can be obtained by increasing the magnitude of the perturbations,the large-scale disturbances are easily detected by the human at low-frequency parts of the image.To address the problem,this thesis proposes an improved attack algorithm WP-AdvGAN,which introduces a positioning module into the original AdvGAN.The positioning module is used for detecting the decision sensitivity of each location to the target model,in order to generate adversarial examples with weighted perturbations.Specifically,locations with high sensitivity in the image often have a greater impact on the decision of the target model,so the perturbations on the corresponding locations are given greater weights.Similarly,for the locations with low decision sensitivity in the image,the corresponding perturbations are given small weights.Experiments denote that the adversarial examples generated by WP-AdvGAN are closer to clean samples,and have higher attack success rates in both targeted and untargeted attacks.