Research On Deepfake Detection With Deep Learning

As deep learning techniques continue to rapidly advance,deep forgery generation technology is also constantly evolving and widely applied in multimedia fields such as virtual reality,film and television production,game modeling,and marketing advertising.However,these technologies can also be used for illegal purposes,such as creating fake political speeches,news reports,promotional materials,or pornography,infringing upon the reputation,portrait,and privacy of others.Therefore,there is an urgent need to develop high-precision,efficient,and lightweight deep forgery detection algorithms.Existing detection methods for fake face images suffer from poor generalization and lack of lightweight models.As for fake face videos,existing detection methods often use imagebased frame-by-frame detection and ignore the temporal information of videos.Therefore,corresponding detection methods for images and videos separately are proposed by this thesis.Firstly,a fake face image detection method based on the edge strip region of the face is proposed by this thesis.As the synthesis of fake faces and real backgrounds often results in defects at the edge regions due to parameter issues,this thesis designs a preprocessing workflow based on the Dlib face detector,Graham convex hull scanning algorithm,and mask series operations,which can quickly and accurately extract the edge regions of the face.The Dlib face detector is used to detect the face and facial landmarks,and then the black-and-white mask of the face region is generated by combining the facial landmarks with the Graham convex hull scanning algorithm.The mask is further processed through image morphology operations(dilation,erosion)and image bitwise operations(bitwise negation,bitwise AND)to extract the required edge strip region of the face.The Efficient Net＿b0＿CA model is proposed by this thesis,which integrates channel attention mechanism into the Efficient Net＿b0 structure.Weighting the feature map of the first MB convolutional block of Efficient Net＿b0 can better extract the features of the edge region of the face.Comparative experiments based on the Face Forensics++dataset demonstrate the effectiveness and advancement of this method.Then,a fake face video detection method based on the edge strip region of the face is proposed by this thesis.More stable edge regions of the face are extracted using the Insightface face detector,and a method based on a three-dimensional convolutional structure is used to detect continuous frames of edge images of the face,effectively extracting spatial and temporal information from consecutive frames.Comparative experiments were conducted on the FaceForensics++ public dataset using the I3 D and R3 D two three-dimensional convolutional models and main-stream algorithms,resulting in higher detection accuracy and more effective extraction of spatiotemporal feature information from continuous video frames.