Research On Video Forgery Detection Algorithms Using Phase Correlation Feature And Deep Learning

Millions of videos are made and uploaded on the internet every day with most of the video's content is not authenticated.Besides that,video content editing software like Video Editor,Adobe Photoshop,Window Movie Maker,and Adobe After Effects also are widely available.They support a lot of methods for editing video content easily,and anyone can edit video content at their willing,even the edited content contrasts with original content.Furthermore,recently with the rapid development of deep learning-based techniques that have created videos with characters' faces replaced by other faces automatically,such tools as Fake App,Faceswap.That leads to seeing the video is no longer believing.In addition,an authentic video gives strong evidence more than an authoritative image in court.Therefore,video forensic proves that video authenticity becomes an urgent requirement today.Recently with the steady development of video editing applications,3-D regions in videos are copied,then pasted to other positions and edited the brightness,geometry,and similar things have been edited easily.These 3-D regions can be small 3-D areas inside consecutive frame sequences or whole successive frame sequences.These have become a popular tampering method used in video tampering,and it is challenging to detect.So,it has been urging us to study for finding algorithms to detect this kind of video forgery.In the first part of this study,we propose a method that uses phase-correlation of frame residual for detecting duplication of 3-D regions and localization in videos.Through many experiments,it is proved that the proposed method efficiently detects duplication of 3-D areas of videos.Besides that,manipulations at frame level conceal or imitate the content in the video,these manipulations are simple skills in editing content of the video.Still,they would create forged videos hard to detect,especially by naked eyes.In addition,manipulations of tamper videos at the frame level were strongly supported by video content editing applications.Anyone can perform deletion,duplication,or insertion of a-frames sequence efficiently by one or two actions on video content editing applications.Through recent researches,deep learning has outstanding results.Notably,the convolutional neural networks(CNNs)has achieved exceptional results in solving many challenging vision problems,such as object detection,self-driving car,visual captioning and especially in large-scale image recognition which has motivated us to research and apply recently efficient CNN models for detecting video forgeries at the frame level(video inter-frame forgeries).In the second part of this study,we proposed a method that applies recent state-of-theart CNN models.These CNN models were trained with more than one million images on the Image Net database,we have fine-tuned then retrained on target dataset for detecting some kinds of video inter-frame forgeries.We have also compared the efficiency of these models with each other to find out which architecture of the CNN model is suitable for detecting video inter-frame forgeries.In particular,these models were not directly retrained from video frames,but they were retrained from the residual or optical flow features between consecutive frames.We have performed many experiments to find out the best feature,which was acquired for proposed methods.Besides,we have also conducted some tests to check the efficiency of transfer learning models trained on the Image Net database for this situation.In the testing stage,the classification score was refined to a confidence score to enhance the effectiveness of the model.Besides manipulated video forgery by humans,recently,millions of videos are uploaded on the internet,many videos of which have been manipulated by fully automated techniques change video content.And the development of that techniques has raised dangerous consequences for individuals and society.Especially,for the last two years,Deep learning-based face replacement techniques in the video have been rapidly developed,primarily,Deepfake techniques with important tools like Face App,Faceswap-GAN,Deep Face Lab,and DFaker,which are used to make videos in which contains face tampering.Those facial video forgeries are hardly distinguished by the naked eye,and they can be made from malicious purposes as pornographic videos of celebrities,politicians,fake news,fake surveillance videos,and policy tensions.So,nowadays,facial video forgery detection has become a hot topic of interest amongst researchers in the world.Recently,there are some suggested methods for detecting Deepfake video;most of them are only based on steganalysis features or learned features on spatial or temporal separately.Features which have relation in spatial and temporal in the video are not exploited.Because a video is a set of consecutive frames in temporal,all of the recent methods have not given good results.Therefore,that is still a significant challenge.Through the state-of-the-art techniques of machine learning showed that 3-D convolution kernels could learn spatial and temporal features at the same time and achieved breakthrough performance.In the last part of this research,we have applied and proposed a method using 3-D convolution kernels to build a deep 3-D convolutional neural network to learn Spatiotemporal features in short consecutive frames sequence to detect Deepfake videos.We have experimented on the two largest and popular Deepfake video datasets as Face Forensics++ and Vid TIMIT and compared the efficiency of the proposed methods with the-state of-the-art methods.Through that,it is proved our proposed method is more efficient and accurate than recent methods.