The Research On Deepfake Videos Detection Of Face Manipulation

In recent years,the face manipulation tools represented by Deepfake had reduced the reliability of information from videos and pictures.Deepfake videos threatening people’s biometric information security in the Internet era.So,the technology of Deepfake video detection has become a very hot research topic.It has brought a great benefit and guarantee for society.The main works of this paper are as follows:(1)A model named Cascaded-Hop for Deepfake videos based on SSL(Successive Subspace Learning,SSL)was proposed.This model equips the ability of learning artifact features from Deepfake videos.To make high-quality frame sample datasets of Deepfake,the repetitive and fuzzy frames were screened out by the image pixel matrix feature in the sampling stage,and the size of facial images was adjusted by the RCAN,(Residual Channel Attention Network,RCAN).Experiments on public datasets show that this model performs well by combining the sampling method.(2)A Deepfake videos detector using auto-adaptive weight for sampling frames was designed and implemented.It is a combination model of Mesonet-D of CNN-based and LSTM(Long Short-term Memory,LSTM)of RNN-based.It can automatically adjust the weight of each sampled frame from the video,making up for the lack of using averaging weight for sampling frames and improving the reliability of the Deepfake videos detection results.The experimental results on the public dataset show that this detector has a nice classification effect.