Research On Interchangeable Deep False Face Detection On The Basis Of Deep Learning

Deep false face has recently become the focus of public concern,with fake face images becoming increasingly realistic and even difficult to recognize by the human eye.Meanwhile,the multimedia forensics community has been developing methods to detect these fake images.Few have explored the important issue of generalization capability of forensic models.With the rapid emergence of new types of deep vacation faces,forensic models are undoubtedly an important research topic to improve their generalization ability of detecting new deep vacation face images while ensuring the accuracy of detection.In this thesis,two methods are proposed for the investigation of the interchangeable deep false face forensics.1.In view of the existing Deepfake detection algorithms,such problems as low accuracy and poor interpretability are common,a neural network model combining the double attention mechanism is proposed,which uses channel attention to capture the abnormal features of false faces and combines the location of spatial attention to focus the abnormal features.To fully learn the contextual semantic information of the abnormal part of the false face,so as to improve the effectiveness and accuracy of face changing detection.In addition,the decision-making area of real and fake faces are shown effectively in the form of thermal diagram,which provides a certain degree of explanation for the face exchange detection model.Experiments on the Face Forensics++ open source dataset show that the detection accuracy of this method outperforms Meso Inception,Capsule-Forensics and Xception Net detection methods.2.For the lack of generalization in the existing deep false face detection algorithms,this paper proposes a deep false face detection algorithm based on Stacking ensembling of features of spatial domain,frequency domain,color space and time domain.The features of spatial domain are head posture features.This thesis makes the difference between the rotation matrix of the center of the face standard fixed-point estimation and the rotation matrix of the face region standard fixed-point estimation,and the next step is to take its standardization and then use SVM classification;The frequency domain features are obtained from the discrete Fourier transform of the image to obtain the magnitude spectrum,further squared to obtain the two-dimensional power spectrum,and then compressed to obtain the one-dimensional power spectrum,and then classified by the SVM classifier.The features of color domain are the residual images of four color components,such as H,S,Cb and Cr,and the three statistics of the co-occurrence matrix are calculated as a feature set and sent to the SVM classifier for classification.The time-domain features are heart rate related features calculated with the Deep Phys model,and the last layer of this model is changed to sigmoid activation to obtain values between [0,1],which can also be classified.The above four features are taken as the basic classifier of Stacking ensembling method.With five-fold Stacking,the final classification result of true and false face can be obtained by the Logistic Regression(LR)model.Experiments on open source datasets such as Face Forensics++,Celeb-DF and Deeper Forensics show that the detection accuracy of the method outperforms detection methods such as integrated CNN and Xception Net.