Research On GAN-generated Face Detection Algorithm Based On Convolutional Neural Networ

With the rapid development of network technology and the increasing popularity of image acquisition devices,colorful digital images have become an important medium for transmitting information.While bringing convenience to people’s lives,it also has potential safety hazards.Therefore,how to protect the integrity and authenticity of digital images is a challenge that digital image forensics will continue to face.In particular,GAN has achieved great success in the image field in recent years,and its generated high-definition images that can be confused with the real have attracted the great attention of experts and scholars in relevant fields.Mining the differences between natural images and GAN generated images,and then accurately and efficiently identifying the forged images generated by GAN is of great significance in the field of digital image forensics.Therefore,this paper deeply analyzes the traces left by the real and fake images in the generation and acquisition process from the perspectives of the spatial domain and the frequency domain,and based on this,two detection algorithms for face images generated by GAN are proposed.The specific research content as follows:(1)Aiming at the problem that most of the existing detection algorithms are only suitable for simple scenes and are not robust to image post-processing operations,we proposed a new framework—two-stream CNN to detect GAN generated fake images,which contains RGB stream and Photo Response Non-Uniformity(PRNU)stream,respectively.In the preprocessing stage of the RGB stream,the use of random erasing enhances the diversity of samples and assists the network to pay more attention to the difffference in GAN fifingerprints in the image content.The PRNU stream’s construction is based on the uniqueness of PRNU features in real images and the robustness of the features to image transformation.The existence of PRNU guides the network to focus on the changes in the image pixel value itself and enhances the generalization performance of the network.Experimental results on multiple datasets show that the proposed method has apparent advantages in accuracy and generalization and is more robust to various image transformations,such as downsampling,JPEG compression,Gaussian noise,and Gaussian blur.(2)For the existing detection algorithms have a certain degree of misjudgment for high-quality generated datasets,and most of them focus on the image spatial domain,ignoring the distribution characteristics in the frequency domain,we propose a detection algorithm based on convolutional neural networks that combine spatial and spectral domains.Given that GAN images leave clearly discernible checkerboard artifacts on the spectrum due to the upsampling operation during the generation process,we design a learnable frequency-domain rate filter kernel and a frequency-domain rate network to fully extract this feature.In order to reduce the influence of the information discarded in the frequency domain,we also design a spatial domain network to learn the differentiated features of the image content itself,and finally fuse the two features to achieve GAN face detection.It has been proved on multiple datasets that the proposed model is superior to the previously proposed algorithm in terms of detection accuracy in high-quality generated datasets and generalization across datasets.In addition,it also has good performance on the local face dataset generated by GAN,which proves that our model has a wider application prospect.