Research On Synthetic Face Detection Technology For Deepfake

Deepfake(synthesis)face technology mainly uses deep learning algorithm to synthesize face images.This technology uses the Generative Adversarial Networks and Variational Auto-Encoder technology to learn the characteristics of real face,which can generate false face images with high authenticity,rich expression,consistent foreground and background style.Different from the traditional image tampering(editing)technology,the emergence of deep synthetic face technology has brought new challenges to the detection and recognition of fake faces.The deepfake face detection technology has also become a research hotspot in the field of computer vision.Therefore,aiming at the problems of insufficient use of semantic information and limited use of modal information in current detection methods,this thesis studies the detection technology of Deepfake face by using the relevant methods of deep learning technology.The main work is as follows:(1)A Deepfake face detection method based on face semantic information Sifd Net is proposed.In view of the insufficient utilization,poor robustness of face features and semantic information when the detection model faces Deepfake faces,Sifd Net uses the semantic information of face components.In the design of the network model,first extract the semantic features of face components through the shallow network,then integrate the full face image features,after that strengthen the features in combination with the CBAM attention mechanism module,make the detection model pays more attention to the forged details of the face task itself.The experimental results show that compared with the mainstream algorithms such as Res Net and Fdft Net,Sifd Net has obvious performance improvement in improving the accuracy of the model.(2)A Deepfake face detection method based on multi feature fusion Mffd Net is proposed.In view of the limited modal information available in the existing models and insufficient attention to feature correlation,Mffd Net introduces the frequency domain information of face image,comprehensively considers the accuracy and time cost performance indicators,and then designs the dual branch model architecture combining traditional statistical features and deep learning.Firstly,the lightweight Sifd Net and the Res Net added with the lightweight module are used to extract the features in image domain and frequency domain respectively,Then feature fusion is carried out before the final classification,so as to improve the attention of the detection model to the correlation between face multimodality and features.The experimental results show that the method in this chapter further improves the performance of the algorithm through the improvement of work 1.Finally,based on the above two algorithms,an interactive system for Deepfake face detection is designed.To sum up,this thesis proposes forward targeted solutions to the problems in the research on synthetic face detection technology for Deepfake,which has a certain reference value for the research of Deepfake face detection technology.