Research On Face Forgery Detection Algorithm Based On Multi-Feature Fusion

Forgery detection as a defense for face recognition systems has developed rapidly in recent years.As the images or videos generated by deep forgery techniques are getting closer and closer to real faces,face forgery detection is facing a great challenge.In order to cope with the complex and variable forgery techniques,the thesis improves the accuracy of forged face detection by fusing two or more features,and the main research contents are as follows:Firstly,a forgery detection algorithm based on the attention mechanism is proposed to address the limitations of single features.The algorithm obtains the local features of the image by adding the attention mechanism and then combines the local features with the global features obtained from the backbone network for face image forgery detection.The global features are selected to be obtained from the backbone network,such as Xception Net or Res Net;and the local features are obtained by adding the attention mechanism on top of the backbone network while introducing the diversity loss function to force the attention map to focus on different regions of the face and discover more artifact features.Finally,experiments are conducted on the Face Forensics++ dataset,and the detection accuracy of the proposed algorithm is improved on different backbone networks.The forgery detection algorithm based on the Res Net-50 backbone network is tested on Face Forensics++c40 up to 94.93% and on Face Forensics++c23 up to 98.36%.The result shows that the proposed algorithm is feasible for detecting forged faces.To address the problem that a single texture feature or depth feature cannot cope well with the current diverse forgery techniques,an algorithm is proposed to fuse the enhanced texture features with the depth features.The proposed algorithm firstly uses the Efficient Net-V2 network to extract shallow features in pixel-level images;secondly,the shallow features are fed into two branches of the deep network and texture enhancement module to obtain artifact features at different levels;and bilinear pooling is used to fuse high-dimensional information and low-dimensional information;finally,the proposed algorithm is tested on FFIW10 K and Face Forensics++ dataset respectively,and the accuracy achieved on both datasets is better,where the test accuracy on Face Forensics++c23 can reach 98.43% and 95.56% on the FFIW10 K dataset,and the effectiveness of the proposed algorithm is verified by conducting experiments on both datasets.In order to make full use of the global and local features in the face region,so that the local features can be used as auxiliary information to guide the global features in the forgery detection process and discover more discriminative difference features;a forgery detection algorithm based on multidimensional feature fusion is proposed.The proposed algorithm consists of two parts: one is the generation of local activation maps by the lightweight network Mobile Net-V3 to obtain local features;and the other is the fusion of texture depth features extracted by Xception Net into global features.The fusion of these two parts yields multiple-dimensional features that are used to detect real and fake faces.Experimental results show that the proposed algorithm is effective in the Face Forensics++ dataset and achieves good performance.The test accuracy on Face Forensics++c23 can reach 98.75%,which indicates that the proposed algorithm has strong classification ability.