With the breakthrough of deep-learning technology in the field of artificial intelligence,deepfake portrait videos appear more and more frequently,such as facial tampering,pornographic video face swapping,changing politicians'faces and making false statements,etc.This kind of deepfakes may pose a threat to societies;therefore,distinguishing deepfake videos from genuine ones has become an urgen issue.Lots of deepfake detection methods are carried out by constructing many data sets with different compression factors.At present,the deepfake detection technology based on deep-learning algorithm is popular,which requires lots of significant time consumption and massive computing power for training classification model.At the same time,the attributes of black box and unexplainability of deep learning networks also plague the researchers in forensic science.In order to solve the problem of authenticity forensics of the deepfake portrait videos,this paper takes portrait videos encoded by H.264/AVC as the research object,and proposes a method based on inter-frame quantization parameter intensity value to detect deepfake portrait videos and real portrait videos.The selection of inter-frame quantization parameter intensity value and the determination of the inter-frame quantization parameter intensity by binary Logistic regression equation are expounded in detail.The experimental results show favorable accuracy and robustness for the deepfake portrait videos synthesized by DeepFaceLab platform.The paper proposed an interpretable detection method for deepfake portrait videos,which is conducive to determine the direction of investigation and confirm the criminal facts.But there are some limitations.Firstly,the experimental samples are not rich enough.Secondly,the introduced method is greatly affected by video compression,which caused limited application scenarios.Thirdly,the analysis efficiency needs to be optimized.
关键词
视频侦查/深度伪造人像/视频帧间关系/量化参数/二元Logistic回归方程
Key words
video investigation/deepfake portrait/relationship between video frames/quantization parameter/binary Logistic regression equation
维普
万方数据