Research On Locally Generated Face Detection And Localization Algorithms

With the widespread application of deep learning,fake faces generated by face2face,face swap or Generative Adversarial Networks(GAN)are spreading in the Internet.Therefore,it is particularly important to study effective face forensics technologies.For the detection of generated faces,the current research mainly focuses on the whole face image is generated.But in some real scenes,such as face image restoration,glasses removal,mask removal,etc,only some small local regions in a generated face image are generated,even very small regions,and most of the remaining regions are natural.Since the generated region may be very small,it may be reduced to a point in a deep convolutional network with multiple pooling layers,or even not at all on the feature map.This may result in poor performance with the globally generated face detection method in locally generated face detection.At present,as far as we know,there is no publicly locally generated face dataset.For detection research,Xception,as a model widely used in global generated face detection,has good performance,but it is still not fully applicable to locally generated face detection in this article.For localization research,RRU-Net is the latest model with good performance for ordinary tampering of non-face images.However,its robustness still needs to be further improved when it is applied to the localization of locally generated faces with attacks.Therefore,this article will focus on the detection and localization of locally generated faces with the following three research works:(1)In order to carry out the research work of locally generated faces,this paper uses Matlab to create binary mask images with different sizes and shapes.Based on the 70,000face images from the FFHQ real face database,we combine the created binary mask images with the original face images to obtain the face images with missing region.And then the missing regions are deep inpainted by the published pluralistic image completion method.The first locally generated face dataset LGGF based on Generative Adversarial Networks(GAN)is constructed.The dataset has a total of 840,000 images.(2)An improved Xception detection model is proposed:(a)Four residual blocks are deleted,parameters are adjusted,and attention mechanism is added;(b)Inception module with dilated convolution is used to obtain multi-scale features;(c)The Feature Pyramid Networks(FPN)is utilized to obtain multi-level features.Experimental results show that the proposed improved Xception model is better than the existing models in terms of the accuracy,robustness and generalization,especially for face images with smaller generation areas.(3)A robust localization model for locally generated faces that combines RRU-Net and the denoising operation module is proposed.In order to improve the robustness of the model,firstly,the denoising operation module is introduced to resist the noise disturbance in the deep network,so as to enhance the learning ability of the model to the locally generated region features.Then,the locally generated loss function(LG_{Io U} loss)is proposed and combines with sample balance loss function(focal loss),which can effectively improve the network attention to the locally generated region in the training process.Finally,a series of ablation and comparison experiments verify the improvement of the proposed model over RRU-Net and the superiority of other existing models.