Research On GAN-generated Images Recognition Algorithm

In recent years,the Generative Adversarial Network(GAN)has achieved great achievements in the field of image generation,and people can easily generate realistic fake images using various GAN models.While people enjoy the convenience brought by GAN models,if such images cannot be accurately detected,GAN-generated images will pose a great threat to social public safety.Therefore,it is imperative to conduct in-depth research on GAN-generated image detection algorithms.The existing models mainly use the classical convolutional neural network to detect the GAN-generated images by extracting the spatial and content information in the image,such as shape,se-mantics and color.However,these methods ignore the difference in global texture information be-tween GAN-generated images and real images,resulting in poor detection accuracy in the face of known GAN models,that is,the training set and test set come from the same GAN structure.In ad-dition,these methods do not make full use of the ”common features” left by GAN-generated images in frequency domain,spatial domain and data statistics,resulting in poor detection generalization of GAN-generated images when the GAN model is unknown,that is,the training set and test set come from different GAN structures.To solve the above problems,this thesis first proposes a new network model: Gram＿Res Net model.The model uses gram module to obtain the global texture features of the GAN-generated im-ages to improve the detection accuracy of GAN-generated images with known structures; Secondly,based on Gram＿Res Net structure,this thesis proposes a new detection model based on multi feature fusion.By fusing multiple domain features,the new model obtains the ”common features” of im-ages generated by different GAN structures,which effectively improves the generalization of image detection generated by unknown GAN models.The specific research contents of this thesis are as follows:(1)For the detection of images generated by known GAN structure,this thesis proposes a de-tection model based on gram module: Gram＿Res Net model.The gram module can extract the unique global texture features of the image generated by the known GAN structure without being limited by the receptive field of the convolution kernel.Therefore,this thesis adds multiple gram modules in different stages of the backbone network,Res Net50,to distinguish the GAN-generated images from the real image.Experiments show that Gram＿Res Net model can effectively extract the global texture features of the image,and the detection accuracy and robustness are greatly improved.(2)On the basis that the detection effect of the images generated by the known GAN structure reaches a high level,this thesis further studies the generalization of the detection model,and proposes detection model of GAN-generated images based on multi feature fusion.The model extracts the features of frequency domain,spatial domain and co-occurrence matrix through multiple branches,and designs early and late fusion mechanisms to fuse the ”common features” of the generated im-age.Finally,the images generated by the unknown GAN structure and the real images are classified according to the ”common features”.In addition,this thesis also generates mixed datasets and data enhancement datasets to ensure that the model can extract ”common features”.Experiments show that the average detection accuracy of the detection model based on multi-feature fusion on the image generated by the unknown GAN is 4.3% higher than the current state-of-the-art model,and the gen-eralization of the model is further improved after training on the mixed dataset,which can be used as a detection model of GAN-generated images.