Deep Fake Video Detection Algorithm Using 3D Soft Biological Features And UV Texture Map

The rapid development of deep learning and digital multimedia technology has made lots of convenience to our daily life and,at the same time,has also brought new challenges to social security.In recent years,the spread of Deep Fake videos on the Internet has become a hot issue.The deepfake technology is able exchange the identities of different characters.A malicious user can obtain a lifelike deepfake video in a few minutes with just a mobile phone.Deep Fake videos can be used to manipulate public opinion,commit fraud,smear,blackmail and other illegal activities.Therefore,an efficient detection algorithm is urgently needed to curb the further spread of this type of misinformation.Most of the existing algorithms only use twodimensional face information.Three-dimensional face information is rarely used for discriminating between the true and false faces.In addition,current approaches rely mostly on image pixel level features rather than biological characteristics,failing to reveal higher level semantic flaws from the deepfake videos.To address these problems,this thesis proposes to exploit techniques like three-dimensional face deformation model,soft biological characteristics and UV texture map.The major contributions are summarized as follows:1?We first introduce the methods of building a three-dimensional face morphing model,including three typical models: the traditional three-dimensional morphable model(3D Morphable Model,3DMM),the BFM model and the Face Warehouse model.We then introduce to use the external parameter matrix of the camera model to estimate the pose information and use the 3DMM model to extract the face shape,texture,and expression parameters.We also introduce that the texture information in the three-dimensional space can be represented by the face UV texture map.The features extracted from the above three parts are the basis for authenticating videos in this thesis.2?We propose a deepfake video detection algorithm based on 3DMM soft biological characteristics.The 3DMM model is employed to extract the single-frame facial attributes from image frames.The specific facial behavior features of a person is learned with an Res Net-34 network and metric learning loss function.We also construct an appearance feature extraction module and establish a reference set based on closed set retrieval.The authenticity of video is determined by comparing the consistency of appearance features and facial behavior features.Experimental results show that the proposed algorithm has better robustness against video compression and higher detection accuracy.3?We propose a two-stream detection algorithm for deepfake video based on UV texture map.The 3D face texture information is used for forgery detection.A two-stream model framework is designed based on the Efficient Net.The face UV texture map generated by a3 DMM model is fed to one of the branch for 3D texture feature extraction,and the RGB image is fed to another branch for 2D color information feature extraction.Finally,the features extracted from different branches are fused to determine the video authenticity.Experimental results show that the algorithm has good accuracies on a variety of databases.