Deepfake Identification Based On Fine-Grained Image Classification And Inter-Frame Similarity In The Area Of Forensics

Deepfake is a technology that uses deep learning algorithm to replace a specific face into a certain video,which can realize identity transfer.If this technology is used for improper purposes will do great harm to society,the country,and individuals.However,it is difficult to use traditional video authenticity forensic methods to detect videos after the replacement of deepfake technology.Also,the law has not yet clearly determined the effectiveness of forensic authentication involving artificial intelligence.Therefore,it is necessary to study how to improve the detection accuracy of deepfake videos and how to Interpreting the detection results of deepfake videos is of great significance for solving the identification problem of audio-video forensics under the background of new technologies.In the practice of audio-video forensic,we often refer to the technical detection methods such as ambiguity,blocking effect and similarity between image regions to comprehensively evaluate the tampering behaviors such as copy paste,MPEG double compression and frame operation.However,for deepfake videos,because the tampering of this kind of video will not leave obvious traces,it is difficult to use traditional methods for forensic identification.Therefore,the emergence of deepfake videos requires the use of new methods,new technologies and new standards in the identification of audio-video materials.The innovation of this paper is that in view of the current situation that it is difficult to identify deepfake videos,a detection algorithm based on fine-grained image classification and using Xception Net structure(FG-X)is proposed.The detection algorithm takes Xception Net as the backbone network and combines the idea of finegrained image classification to enhance the data of the original image by means of image feature information enhancement and image feature information erasure,then convolution operation and bilinear attention pooling are used to calculate the features extracted for the first time to obtain the feature matrix.The algorithm has the characteristics of high accuracy and good generalization,and its interpretability can be enhanced by feature visualization.In the experimental analysis,firstly,the three network structures of Xception Net,Inception Net V4 and Dense Net are compared.It is found that the detection accuracy of Xception Net is higher than that of the other two models.Then Xception Net is optimized based on fine-grained image classification.The results show that the accuracy in and out of the dataset reaches 99.5% and 96.9%respectively,which is far higher than that of the existing detection models,indicating that the algorithm can effectively improve the accuracy of deepfake video detection.In addition,according to the relevant industry standards and technical specifications of judicial expertise,this paper also puts forward the deepfake video identification model.Specifically,firstly,this paper puts forward a new research field of deepfake video forensic.The research conclusion of this paper shows that deepfake technology will bring new opportunities and challenges to the field of forensic.Secondly,this paper analyzes the similarities and differences of existing detection algorithms at home and abroad.In order to make the detection algorithm better target the face part and apply fine-grained image classification to the field of deepfake detection,a deepfake video detection algorithm based on FG-X is proposed.Experiments show that its detection accuracy of deepfake video is improved,which not only expands the application scope of fine-grained image classification,Moreover,it makes the deepfake video detection technology develop further.Thirdly,although the above detection algorithm can improve the detection accuracy of deepfake videos,it still fails to solve the problem of lack of interpretability,which makes it difficult to be used in forensic practice.Therefore,combined with the relevant norms of judicial expertise,this paper also discusses the details of the inspection material requirements and inspection process of the authenticity identification of deepfake video in judicial expertise,puts forward the dilemma of applying the detection conclusion of deepfake video to the judicial expertise opinion and the corresponding countermeasures,and verifies the application value of the deepfake video detection algorithm based on FG-X in judicial expertise.Finally,the shortcomings caused by time and hardware in this paper are summarized,and the application of deepfake video detection in practical identification is prospected.In short,deepfake video has the characteristics of less tampering traces,difficult identification and lack of relevant identification specifications.Therefore,there will be difficulties in technology and specifications in identification.If we can improve the accuracy and interpretability according to the existing deepfake detection algorithm and the requirements of judicial expertise,combined with the relevant norms of judicial expertise,the application of deepfake video identification in judicial expertise practice is just around the corner.