Research On Generalization Of Generated Image Forensics Model Based On Deep Learning

The advent of the intelligent era has accelerated the development of image gen-eration technologies,especially the rapid development of image generation algorithms represented by Generative Adversarial Network.At present,image generation technol-ogy can generate high-quality images that are highly similar to images taken with a camera in the real world,so that the human eye is hardly distinguishable.In addition,due to the rapid development of the Internet,the threat of generated images has become increasingly apparent.When these realistic generated images are abused or maliciously confusing audiovisual,it will have a serious adverse effect on the credibility of net-work information.In the field of image forensics research aimed at this threat,there are already some algorithms for detecting generated images,but there are still some unresolved problems.The main work of this thesis includes:(1)An algorithm for improving the generalization of forensic models is proposed.This work research found that most of the currently generated image forensics algo-rithms have good detection effect only for a specific dataset,and the detection of other datasets has the problem of weak generalization ability.Based on this observation,this work proposes to use an image preprocessing method to improve the generalization of the forensic model.On the one hand,this image preprocessing can be used to destroy or suppress the low-level forensic clues of generated images,so that the model does not pay attention to the unstable image features of the generated images.On the other hand,the image preprocessing can improve the statistical feature similarity at the pixel level between the real image and the generated image,which can increase the difficulty of model training and force the forensic model to learn more features related to the mean-ing of the generative algorithm.Experimental results verify the effectiveness of this image preprocessing method,which can help forensic models improve generalization performance.(2)A forensic algorithm for generated images of multiple types is proposed.This work research found that the existing forensic algorithms for generating images are mostly a binary classification architecture,and the detection performance of new types of generated images is poor.This work regards the detection task of generating images as a multi-class classification task,and proposes a multi-class classification model based on deep metric learning,and increases the scalability of the model.This work considers the forensics task as a multi-classification task,which can not only divide the image into two types of real and fake images,but also can classify fake images more finely.In order to cope with the phenomenon of new types of generated images,this work established a library of generated image templates,and fine-tuned the model to further improve the detection performance of the forensic model.Finally,the experimental results show the superiority of the method architecture.In summary,this thesis has conducted in-depth analysis and research on the ex-isting image forensics algorithms,and points out the problems of the current forensic models.Based on deep learning and using generated images as research objects,this thesis designs and researches on generated image forensics algorithms.The algorithms proposed by this thesis can effectively solve some existing problems in the field of foren-sics.The research has made useful explorations and valuable reference contributions.