Research On Forgery Face Detection

The forgery faces are generated by manipulating the faces in images or videos with a specific purpose including identity replacement,face expression and attributes manipulation,etc.The rapid development of face manipulation technology makes the manufacture of forgery face more accessible than before.For example,the identity,expression,and attribute of face can be easily manipulated by using the existing tools.Meanwhile,advanced techniques make it extremely difficult for human to distinguish between real and forgery face,which greatly reduces the credibility of digital media and brings many risks to various fields such as politics,economy,and society.This has attracted widespread attention from society,and there is an urgent need for the forgery face detection algorithms.Currently,taking fake face detection as a binary classification problem,the forgery face detector based on deep convolutional neural network can achieve satisfactory detection performance.However,there is still room for the improvement in terms of detection accuracy,interpretability,etc.Overall,developing algorithms for forgery face detection has important academic and practical value.The main contributions of this paper are as follows:1.We propose a novel attention-based data augmentation algorithm,Representative Forgery Mining(RFM),to improve the detection ability of the model on forgery faces during the training process.Existing forgery face detection models have the problem of overfitting facial features and poor generalization performance.The proposed data augmentation algorithm efficiently and specifically erases the overfitting feature regions in images using the backpropagation gradients of the model during the training process.The experimental results show that the algorithm can effectively improve the model’s attention to forgery faces and significantly enhance the detection ability of the model on forgery faces when combined with multiple model structures.2.We propose a spatio-temporal forgery attention network(STFAN)based on temporal-spatial differences for forgery face video detection.Existing networks have the problem of insufficient utilization of temporal information.The proposed algorithm uses two modules,spatial attention and temporal attention,to extract high-frequency forgery features from the spatial and temporal domains of the video,respectively,to generate attention maps.The attention maps guide the network to focus on the forgery traces in the video to improve detection performance while enhancing the interpretability of the detection results.The experimental results show that the algorithm effectively improves the detection performance of forgery faces in videos.