Research On GAN-generated Face Image Forensics

With the rapid development of deep learning,people can easily synthesize extremely realistic face images.In particular,the Generative Adversarial Network(GAN)proposed in recent years,which can synthesize high-resolution face images,has achieved the effect of deceiving people easily.Since face images contain rich semantic information,when generated face images are maliciously used by criminals,it will cause huge ethical,moral and legal problems.Therefore,GAN-generated face forensics has become a research hotspot in the field of forensics.Different from other face synthesis methods,GAN-generated face does not exist in real life,so it is a completely fake image.Because the GAN-generated face image itself is a fake image and there is no artificial manipulation operation,so the traditional forensics methods based on tampering traces are no longer useful.To adress this problem,this paper designs corresponding forensics algorithm s by analyzing the different characteristics of the GAN-generated face images and the real face images due to different acquisition processes.The main work and innovations of the paper are summarized as follows:First,in response to the forensics problem of real face images and GAN-generated face images,a algorithm based on the interpolation characteristics of the color filter array(CFA)is proposed.Generally speaking,only a single sensor is used in the imaging process of real natural images for cost-saving purposes.Therefore,each pixel has only one color information and the other two color information must be obtained through CFA interpolation.Because of the effect of CFA interpolation,there is a strong correlation between the RGB channels of natural images,while GAN-generated face images are obtained through continuous network training process,there is no such correlation between the three color channels.Based on this,this paper designs a forensic method of manually extracting color channel discriminative features.This method uses wavelet transform to decompose the three color channels of the images,and then extract the statistical features of the mean,standard deviation,skewness and kurtosis of the high-frequency sub-bands of each color channel,and the corresponding correlation coefficient between the RG,RB and GB channels of wavelet sub-bands.Finally,the extracted hybird features are put to the support vector machine(SVM)for classification.Results show that our proposed method can achieve98.10% accuracy.Compared with methods based on deep learning,our proposed method is simpler and lower in cost.Second,in response to the forensic problem of different GAN-generated face images,a algorithm based on sensor pattern noise and texture characteristics is proposed.Usually,real images often leave specific sensor pattern noise due to different sensor discrepancy during the imaging process,which is also regarded as the fingerprint of the image.For GAN-generated images,the synthesis process is through the constant training of the network.Although there is no such image sensor in this process,operations such as convolution in the network actually play a similar role to the image sensor.Therefore,different GAN models will also leave specific “sensor pattern noise” fingerprints in the generated image.In addition,images generated by different GAN models also show different texture characteristics.Based on this,this paper extracts the local binary pattern(LBP)feature o f the image and the subtractive pixel adjacency matrix(SPAM)feature of the image sensor patte rn noise respectively.Finally,we use SVM classifier to classify it.The results show that the accuracy of the proposed method for the face images generated by different GAN models is above90%.It indicates that this method can be used as an effici ent forensic algorithm for GAN-generated faces.