Research On Deep Learning Based General Facial Image Forensics

With the rapid development of computer graphics and artificial intelligence,people can easily generate and edit facial images,and these resultant facial images have scarily-real visual qualities.The advanced face image generation and tampering techniques bring great pleasure and convenience to people's life,yet also bring serious challenges to the authenticity of face images.Face images carry personal identity information.when face images are maliciously manipulated and used,they might lead to serious threat of public confidence.Therefore,face image forensics has recently been becoming a hot topic in the field of image forensics.There are two common attacks torwards face images: one is to generate artifical face images which do not exist in real world by means of various techniques such as adversarial generative network(GAN),and the other is to manipulate face images that captured naturally,which includes face swapping,face retouching and expression transfer.This thesis researches on general face image forensics,which is devoted to address the following two issues.First,to simultaneously detect face image forgeries which might be generated or tampered by multiple techniques.Second,to detect face images which might be obtained from unknown techniques.The main works and contributions are summarized as follows:Firstly,Since most existing face image forensics works are only suitable for the detection of single face manipulation technique for binary detection,a general face forensic approach is proposed to detect those facial images which might be generated or tampered by mulptiple techniques.Actually,the tamering artifacts left by various face image manipulations are not the same,and thus the traditional hand-crafted features are not suitable for universal face image forensics.We design a light-weight convolutional neural network(CNN),which adopts the network architecture made up of deep separable convolution layer,1 × 1 convolution residual structure and batch normalization layer.Compared with the existing works,the proposed approach has the advantages of simple structure and fast convergence.It can detect face images manipulated by various technologies,and it is also robust against different types of post-processing attacks.Secondly,since face image tampering techniques are fast developing themselves,we propose a general face image forensics approach to detect those face images which suffer from unknown face image manipulation techniques.Usually,deep network models are only able to detect those face images by known face image manipulations which have been covered by the training set.The proposed approach is based on the Efficient Net model,and it designs a convolution layer with fixed weights,whichs contain three filter kernels,to suppress image semantic content and highlight image micro-structures,so as to improve the generalization capability of deep neural network.The experimental results show that the proposed approach can effectively improve the generalization capability of the forensic model,and can detect those face images which suffer from unkonwn face image manipulation techniques.