High Fidelity Face Swapping And Multi-Information Domain Forgery Detection

In recent years,with the rapid development of deep learning,face forgery techniques represented by face swapping have become rapidly popular on the Internet and have received wide attention in the fields of computer vision and information security.Face swapping is to replace the identity of the target face with that of the source face,while other attributes remain unchanged.This technology has a wide range of applications in film and television production and popular entertainment.The current face swapping technology is in the development stage,and there is still room for further improvement in the naturalness of its generation.On the other hand,this kind of face forgery technology can be easily used by lawless elements,which poses a great threat to national security and social stability.Therefore,it is crucial to propose effective face forgery detection methods.The forgery detection methods at this stage have an excellent performance within the dataset but poor generalization capability.In order to promote the positive application of face swapping and reduce the negative effects of such technologies,this thesis has improved the above problems and the main work includes:(1)Most of the existing face swapping methods only focus on the maintenance of identity information and ignore the restoration of expressions,so they are prone to expression distortion problems.To solve this problem,this thesis improves based on adaptive embedding integration network and proposes a face swapping algorithm toward expression high fidelity,firstly,adding a face reenactment module to reduce the influence of irrelevant attributes in the source image by synchronizing the expressions and poses of the source and target faces;secondly,using a new attribute matching loss to train the face swapping module to improve the consistency of the generated results and the target face expressions.Experimental results show that the proposed face swapping algorithm has good expression restoration ability,while visually reducing the appearance of artifacts and fusion boundaries.(2)In face forgery detection,methods based on a single information domain tend to perform well within the dataset but have poor generalization across datasets.In response to this problem,this thesis proposes a multi-information domain face forgery detection algorithm based on the spatial and frequency domains of images,using multi-level frequency decomposition and a two-stream network to achieve complementary spatial domain and frequency domain information,and adding a self-attention-based data enhancement method to deeply mine the representative forgery information of faces to further improve the accuracy of detection.The experimental results show that the proposed face forgery detection algorithm has a good detection effect both inside and outside the dataset,and is more generalizable compared with other algorithms.Figure [26] Table [7] Reference [81]...