| Deepfake videos with manipulated visual content generated by artificial intelligence technologies are becoming increasingly common on social media and online platforms.These videos not only spread false information but also manipulate public opinion and perpetuate disinformation campaigns,posing a serious threat to individuals,organizations,and entire societies.Therefore,detecting Deepfake videos to reduce the negative effects of deep forgery technology has become an urgent problem to be solved.The main innovations of this paper are as follows.(1)Research on detection of compressed Deepfake videos in social networksIn response to the problem of compressed artifacts interfering with forensic clues in compressed videos on social networks,this paper explores frame-level and time-level perspectives to improve the performance of detecting compressed Deepfake videos.Specifically,this paper analyzes the compression noise brought by compression to facial frames and the differences between true and fake videos on a time-level basis.Based on these analyses,a dual-branch detection method for compressed Deepfake videos based on both frame-level and time-level information is proposed.On one hand,the frame-level branch reduces the interference of compression artifacts on the manipulation trace extraction process through network pruning.On the other hand,the time-level branch detects the differences between true and fake videos on a time-level basis by extracting time-related features.This paper performs experiments on compressed datasets and cross-compression datasets.The experimental results show that the proposed detection scheme greatly improves the detection performance of compressed Deepfake videos.(2)Research on Deepfake video detection with high-visual-qualityTo address the issue of weakened forensic clues in high-quality Deepfake videos,from a new perspective,this paper transforms the problem of Deepfake detection into a frame prediction problem to detect high-quality Deepfake videos and provide an explanation and analysis of unique forensic clues for high-quality Deepfake video detection.Specifically,this paper proposes a frame prediction-based method for detecting high-quality Deepfake videos.Frames are adaptively divided into reference source frames and reference target frames,and facial frames are represented as vectors using a deep neural network.Then,based on the vector representation of the reference source frame,the vector representation of the predicted target frame is obtained by predicting the vector representation of the target frame.The detection of high-quality Deepfake videos is achieved by comparing the correlation between the facial representation vectors of the predicted target frame and the reference target frame.By adopting information-theoretic analysis,it is found that there are differences in mutual information between high-quality true and fake videos in consecutive frames,which confirms that the proposed frame prediction method can improve the detection performance of high-quality Deepfake videos.Performance evaluation is conducted on multiple high-quality Deepfake video datasets,and experimental results demonstrate the effectiveness of the proposed frame prediction-based detection scheme for detecting highquality Deepfake videos.(3)Research on detection technology for cross-domain Deepfake videosTo address the issue of weak generalization performance in cross-domain Deepfake detection due to continuous updates of Deepfake videos,this paper proposes a method that masks and reconstructs different regions of real faces to train models to learn consistent features and avoid overfitting to specific facial attributes and manipulation traces,providing a new clue for Deepfake detection.Specifically,this paper proposes a cross-domain Deepfake detection method based on masking and reconstructing facial regions.A selfsupervised model is used to learn consistency features of facial regions and build a consistency model.The differences between true and fake videos are detected by fine-tuning the consistency model and fusing multi-frame information,while the differences between reconstructed faces are enlarged through mapping.In this paper,the performance was evaluated in multiple cross-domain detection scenarios,and the experimental results show that the proposed detection method can significantly improve the detection performance in multiple cross-domain detection scenarios.In summary,this paper conducts key technical research on cross-domain Deepfake video detection in multiple scenarios focusing on compressed Deepfake videos,weakened forensic clues,and continuous updates of Deepfake videos.By solving the identified problems and advancing key technologies,the research results not only improve the detection performance of Deepfake videos but also introduce new perspectives for Deepfake video detection in a broader field. |
