Research On GAN-Based Synthetic Image Discrimination

Generative adversarial networks(GANs)are trained to learn a particular thing from a large amount of data,to summarise its distribution at the mathematical level and construct a reasonable mapping function to solve real-world problems.This kind of network structure has made a big difference in the field of deep learning,both in computer vision(CV)and natural language processing(NLP),where many excellent GAN structures have emerged.For example,Pro GAN,which generates image datasets and reduces the cost of manual collection and annotation;Ti VGAN,which converts text into images and provides a new form of expression for art;Star GAN,which edits photos for online hair styling,online clothing fitting,etc.;and DSD-GAN,which performs photo restoration and improves the resolution of photos.competing design features in a zero-sum game framework across all types of image processing research.While GAN solves many complex problems,it also brings some concerns from the other side.Because of the wide range of operations that can be performed on images and the extreme realism of the generated effects,this poses a great threat to the trustworthiness and security of traditional visual information.However,existing identification methods are either developed for a specific GAN,which is not universally applicable,or use large amounts of data for identification through extensive training,which requires extremely high computing power of the device and takes a long time.In this study,by analyzing the characteristics of existing GAN models,exploring the commonality of different GAN models in training and generation,and finding the common features in the generation process of existing GAN models for image generation,we propose a power spectrum-based synthetic image identification method,and conduct identification tests on the images generated by several common GAN models and compare them with existing methods.The following work is done:(1)Research on the application of GAN-based synthetic images in data enhancement.We use Star GAN based on domain transformation and Style GAN based on style blending to process a set of real frontal face images,and generate a set of synthetic face images based on light,expression,angle and other conditional transformations or slight occlusion,and use Res Net-based binary classification network to train the synthetic images for face classification,and then use the real It is demonstrated that the classification network trained with synthetic images can achieve accurate face recognition,and the simulated experimental results verify that synthetic images are not essentially different for neural networks in the traditional spatial domain.(2)Power spectrum-based GAN generative image feature point study.Firstly,we analyze the underlying structure and generation principle of the generative GAN and find that the GAN model includes an upsampling layer,which is crucial to the image generation process,and the best upsampling method is transposed convolution and nearest neighbor interpolation,which leads to many abnormal high frequency information in the image and cannot be filtered out without compromising the image quality.Secondly power spectrum calculation of the image can highlight the high frequency information in it very well.Then,the power spectrum data of the synthetic images generated by several common GAN models are calculated and compared.The experiments prove that the synthetic images generated by common GAN models have similar abnormal high-frequency distribution,and the high-frequency information generated by each GAN model has its own unique characteristics.(3)Research on synthetic image discrimination based on power spectrum.Firstly,the average power spectrum is calculated using many real images as a set of benchmark data,and the normal high-frequency signal distribution features are taken as the set of data.Secondly,the baseline feature distribution is mathematically modeled with the abnormal high-frequency signal distribution of the synthetic images by using the Singer function with the weight coefficient matrix,and the high-frequency distribution features of the images generated by each GAN model are extracted using the calculation of the spatial difference degree for the purpose of maximizing the differences between the synthetic images generated by different GAN models.Finally,we use Res Net as a classification network to conduct multi-classification discrimination training,and conduct comparison experiments with existing time-domain discrimination methods and frequency-domain discrimination methods to prove that the power spectrum-based discrimination method has lower time complexity and higher discrimination accuracy,and the experimental results prove that this study can not only achieve efficient discrimination of synthetic images,but also can accurately trace the source of synthetic image generation models.