| With the development of deep learning,“AI face-swapping” techniques such as deepfake are widely spread on the Internet,which uses deep learning algorithms to create and synthesize visual content to achieve video simulation and forgery.The rapid development of Deepfake technology has brought huge security risks to our daily life,accompanied by potential negative impacts in the future.Due to the high quality of the current common deepfake video,it is difficult to be distinguished by the naked eye,and the deep learning model is also widely used in various deepfake generation methods,which makes the detection of deepfake more difficult.Although some existing detection methods have achieved reasonable detection performance in statistical evaluation indicators,there is almost no relevant research on deepfake forensics traces.In this context,according to the research idea of process analysis → artifact forensics → tamper detection,this thesis starts from the perspective of image forensics,aiming at the problem that the current deepfake detection method has weak interpretability and low detection accuracy.By analyzing the operation steps of the deepfake generation process,the common tampering positions and forms in the generation process are grasped in detail,and the traces with obvious differences between real and fake images are found.Combined with the idea of deep learning,the deepfake detection method based on artifact recognition is explored.The main work and innovation of this thesis are reflected in the following two aspects:(1)Aiming at the problem that the face replacement operation in the process of deepfake generation leads to the change of the underlying noise of the image,a deepfake detection method based on the analysis of the underlying noise difference between real and false images is proposed,which realizes the high-precision detection of deepfake.Aiming at the problem that the underlying noise of the picture is difficult to obtain,a novel face-background strategy based on the Siamese network is proposed to train the underlying noise extractor to accurately find the underlying noise traces in deepfake pictures.Aiming at the research of face-background noise trace correlation,a similarity matrix module is proposed to analyze the noise similarity coefficient of the face area and background area in the picture.Finally,a classification module is designed to classify the similarity coefficient of noise to judge the authenticity of the picture.The experimental results show that the accuracy of this method is more than 99.1 % within multiple datasets and 88.9 % across datasets,which exceeds most of the current algorithms and proves the correctness of the theoretical model.(2)Aiming at the problem that the existing deepfake images are high quality and difficult to obtain evidence directly,a novel network structure is proposed to amplify the change of contour information caused by deepfake generation,which can obtain better detection performance.Aiming at the problem that the fusion trace is not obvious in the generation process,an image reconstruction network is designed to enlarge the image contour with obvious differences between real and fake images,and the image fusion trace is clearly found.Aiming at the problem that the image reconstruction process is difficult to supervise training,a new supervised learning method,selective supervision,is proposed to supervise the learning process of the image reconstruction network.This method takes the reconstructed image as a sample,selects the corresponding desired image as a label,and uses the twin network to shorten the feature distance between the reconstructed image and the desired image.Finally,the Vision Transformer is used to predict the authenticity of the reconstructed image.This method achieves good detection results on multiple recognized data sets.In addition,the article clearly shows the edge contour traces after reconstruction and amplification,which proves the effectiveness and robustness of the model.The two core tasks proposed in this thesis analyze the artifacts in deepfake video from the perspective of forensics,train and test on the recognized data set of deepfake.The experimental results show that both model frameworks can improve the accuracy and generalization of deepfake detection. |
