Research On Face Tampering Detection Method Based On Convolutional Neural Network

In recent years,with the upgrading of computer hardware equipment and the continuous development of deep learning,the tampering of image and video has become more and more easy.Especially the identity-marking features of human faces,if tampered,will cause serious social problems.Therefore,effective face tampering detection algorithms are particularly important.Many researchers have done related research in the field of face tampering in recent years.However,as there are more and more methods of face tampering,and the resulting tampered face becomes more and more realistic,the previous methods basically fail.In response to the above problems,based on neural network,this paper will improve the algorithm of face tampering detection from the two aspects of image augmentation and noise filtering.For the pretty lifelike characteristics of face tampering images,this paper designs a combined image augmentation method that makes the generated images diversified while effectively reducing image degradation and retaining the difference between real and fake facial images.At the same time,this article also proposes a training strategy for mixed data streams,which can make image augmentation play its advantages while reducing the impact of data fluctuations caused by image augmentation on the model,and improve the generalization ability of the model to a certain extent.The combination of these two methods can achieve 93.25% and 93.91% accuracy on the Face Forensics++ and Celeb-DF data sets,respectively.For the problem of poor detection performance on face images generated by unknown face tampering methods,this paper also proposes a new method.The tampered face image inevitably contains the fusion of multiple face images.Therefore,compared with the normal face image,the tampered face image must contain more abnormal noise.Accordingly,this paper uses SRM filter to detect the noise of a face image,and then designs a two-stream network to let the model learn the difference between the RGB image and the filtered image at the same time.In addition,this paper also designs a two-stream feature fusion weighting method and a hierarchical feature fusion method to more fully integrate the features of the two data streams,and make the output result more discriminative.The accuracy of this method on the Face Forensics++ and Celeb-DF datasets reached 96.02% and 97.05%,respectively.