| Deep learning has become one of the hottest technologies in many research fields because of its excellent ability to express features,and it continuously contributes reliable technical solutions to complex practical problems in different research fields.In recent years,deep forgery is a typical representative,and the wide dissemination of image forgery technology is becoming more and more complicated,which brings new problems and challenges to maintaining the security of cyberspace.In view of the unstable factors of malicious image forging in cyberspace,many scholars have proposed solutions to this problem.However,there are two key problem in that research of counterfeit image forensics at present.One is the lack of practical and effective extraction method for distinguishing the distinguishing features of counterfeit images in the face of complex and changeable real counterfeit scenes,the other is that the method of face forgery image detection has insufficient generalization ability and poor practicability,which leads to a very passive situation in the confrontation with forgery technology.It is not conducive to the sustainable and stable development of cyberspace.In this paper,based on the deep learning method,we propose a face forgery detection method based on image reconstruction and an improved detection method based on depth metric learning.The experiment shows that,on the basis of ensuring the detection precision of forged face image,a two-stage detection model is constructed by reconstructing the image and learning the similarity,so as to enhance the model’s ability to capture the discriminative features,improve the generalization performance of the detection model.The main research results are as follows:(1)The method of face forgery detection based on image reconstruction is proposed.In that method,according to the flow of face forgery,the solution to the problem of image discrimination of face forgery is divided into two stages:enlarging the forgery feature and capturing the forgery feature and establishing a stable and effective mapping relationship between the feature and the image authenticity.Through modularization design,corresponding capability modules are designed for the two stages of tasks,and a face forgery detection model based on deep learning is implemented.In that process flow of the forged image,the reason of the frequency component fission caused by the up-sampling operation is analyzed,the conclusion is drawn that the new frequency component is inherited from the original component,and then the forged trace is further enlarged by reproducing the forged process,the traces of falsification in the frequency domain are then captured by residual learning.The feature information in the spatial domain and the frequency domain is comprehensively utilized,the stability of the model is strengthened,the composite loss function is constructed and different learning rate adjustment strategies are used to improve the performance of the method.The residual is actively constructed to suppress the high-level semantic information to reduce the negative impact of non-discriminative features and to enhance the robustness of the model.The experimental results show that this method has good performance in face forgery detection,and the model can still show excellent versatility and stability in the face of unknown detection samples.(2)An improved detection method based on deep metric learning is proposed.Based on the method of face forgery detection based on image reconstruction,we continue to explore ways to improve the performance of model detection,and further propose an improved detection algorithm based on depth metric learning.Aiming at the classification module,the algorithm flow is optimized to reduce the influence of human factors on the task of face forgery detection.At the same time,the twin network idea is adopted to design a twin classifier suitable for the face forgery detection task,and class enhance contrast loss function is proposed to the contrast loss function between classes,so that the twin classifier can serve the object of the image forgery detection task better.Experiments show that the improved detection method further improves the generalization performance of the model on the basis of providing good detection precision,and proves the validity of the method of face forgery detection based on image reconstruction.It provides a new solution for face forgery detection. |
