Facial Deep Forgery Detection Research Via Deep Learning

With the widespread application of deep learning,many synthesized pictures and videos are widely disseminated on the Internet.Deep learning-based facial deep forgery techniques whose products make it impossible to distinguish the real and the fake.The field of deep learning continues to emerge with new forgery methods,making the trend of facial deep forgery proliferate.Since new forgery techniques can reduce the effectiveness of the previous detection methods or even render them invalid,which brings new challenges to forgery detection research.The study of facial deep forgery detection needs to move forward in its development.In addition,the unrestricted proliferation of facial deep forgery techniques could jeopardize the security of personal and national data information.Therefore,it is crucial to investigate the detection and identification of facial deep forgeries by training deep learning-based models on real and fake video data and then using the models to detect and discriminate them.This dissertation conducts the following research works on the task of facial deep forgery detection.Starting from the study of image frame-based facial deep forgery detection,this dissertation proposes a detection model based on the combination of image gradient and deep neural network,which transforms the facial deep forgery detection problem into the feature recognition and analysis of one frame in the video.First,it captures a frame from the video,crops it and keeps the face part to reduce the data volume and improve the computational efficiency.Second,it uses the image gradient operator to process the facial image and combines the deep neural network to perform supervised training on the features of the extracted image frame to achieve forgery detection.Extensive experiments on different facial deep forgery datasets show that the image gradient-based approach can effectively detect facial deep forgery and achieve better detection and discrimination performance.Starting from the study of facial deep forgery detection based on consecutive video frames,this work finds that the optical flow imaging of the forged facial video is truncated between consecutive frames after intensive optical flow processing,and this phenomenon is more obvious when the facial action amplitude is large.This dissertation proposes a facial deep forgery detection model based on these findings,i.e.,a unique facial double triangle region is used to assist in capturing facial actions,to achieve the extraction of video inter-frame optical flow feature data,and then to use these features to train the detection model on the dataset.Experimental results verify the relationship between dense optical flow truncation and facial actions,and extensive experimental evaluations show that the proposed method is effective for facial deep forgery detection,achieving better detection results.Figure[27]Table[8]Reference[84]...