| In recent years,with the rapid development of deep learning technology,significant progress has been made in the field of image editing.The "deepfake" creation technology with the face as the main object of tampering has been able to forge fake videos that are enough to deceive human eyes.These deepfake creation techniques can be used to create photorealistic multimedia content that can greatly benefit industries such as film production,photography,video games,and virtual reality.But it may also be used for malicious purposes,such as making fake pornographic videos for extortion,or creating fake news to manipulate public opinion,etc.Therefore,it is imperative to study the detection technology for deepfake videos.Although current researchers have also proposed a variety of detection schemes,but none of these schemes catches the traces of forgery,and the detection effect is not ideal.This thesis finds that the current fake videos will basically leave some visual artifacts,which are mainly spatial artifacts and temporal artifacts.The spatial artifacts are mainly manifested as some local distortions in the RGB domain.The temporal artifacts are mainly manifested as inconsistencies between the front and rear frames,and instability in the face regions.These visual artifacts are the key information for identifying videos,and the deepfake detection models should pay utmost attention to these artifacts.Therefore,the thesis designs two deepfake video detection schemes for the perspective of capturing visual artifacts.The main contributions are summarized as follows:1.A deepfake video detection method based on 3D attentional convolutional network is proposed for temporal artifacts.In order to make the detection model better captures the spatio-temporal inconsistency information in the video,the thesis chooses a more suitable 3D convolutional network as the backbone network of the model.At the same time,in order to match the 3D convolutional network,we design a novel video data preprocessing method and a data augmentation method.In addition,we extend CBAM and proposes a 3D spatio-temporal attention module and a channel attention module to further enhance the model’s ability to perceive subtle features.2.A deepfake video detection method based on fine-grained feature learning is proposed for spatial artifacts.The thesis considers an image processing operation that forces the model to focus on subtle features and enhances its learning ability for fine-grained features by segmenting and breaking up the input images.At the same time,in order to match different granularities of segmentation and scattering,we improve the network structure by introducing multiple branch classifier networks at different levels of the backbone network.We use different classifier networks to match input images with different granularity for training.Finally,the features of different levels of the backbone network are combined to comprehensively determine the authenticity of the input picture.3.A large number of comparative experiments and ablation experiments are designed on two mainstream deepfake public datasets,FF++ and CelebDF.These experiments demonstrate the superior performance of our method over current state-of-the-art detection methods. |
