Research On Face Anti-spoofing Detection Based On Convolutional Neural Network

The rapid development of face recognition technology makes human life more intelligent.But at the same time,the cases of face recognition system being attacked and cracked cause the security problem of face recognition to be paid more and more attention.Among many security problems,face spoofing attack is one of the most noticeable problems in face recognition system.Imposters use electronic devices and other ways to forge the user's facial information and break the face recognition system.Therefore,the accurate detection of the spoofing face has great significance to improve the security of the system.From the original manual features to the current features based on convolutional neural network,the core idea of face anti-spoofing detection is to find the different clues between the real face and the spoofing face,and then make a binary classification by means of machine learning.Training a face anti-spoofing detection model with high security performance based on this traditional method,a large number of datasets with complete contents and all kinds of different attack types are required.Since it is very difficult to collect datasets for face anti-spoofing detection,and imposters can always create new modes of attack against the system,it is almost an impossible task to create an anti-spoofing dataset containing all the types of attack.Therefore,the generalization and robustness of the model is highly required to adapt various attacks.Furthermore,it can be found that the phenotypic characteristics of the attacking samples are greatly different in the task of face anti-spoofing detection.Because the different textures of spoofing mediators will have different feature representation and resulting in a sparse distribution of the spoofing eigenvectors in the multidimensional space.Therefore,in the process of feature extraction or classification,the sparsity of negative sample's distribution should be kept as far as possible.The sparse space is helpful to express abundant attributes of spoofing face.However,the conventional classification methods ignore this peculiarity,which lead to insufficient generalization of the model.Based on the above observations and reflections,we propose two face antispoofing detection methods aimed at improving the generalization of convolutional neural network:(1)This paper proposes a face anti-spoofing detection method based on neural network with buffer layer,which aims at improving the robustness of the model for the purpose of lifelong learning,to make the model have a better recognition effect when suffering the constantly changing attacks.First of all,different from the general method of extracting the whole facial information as the input of the model,this method only integrates the features of the facial significance areas,so that the model can effectively suppress the background interference.Furthermore,in order to improve the generalization of the network,the proposed buffer layer architecture can realize incremental lifelong learning on the premise of training only a full connection layer.To make the model converge effectively,this paper also propose a Mag Loss function based on Maximum Mean Discrepancy(MMD),which effectively restricts the difference of distributions and inhibits catastrophic forgetting.Experimental results show that the proposed method can effectively improve the generalization of the model while keeping the training speed.(2)This paper also proposes a face anti-spoofing detection method based on Dense Sparse Inverse loss function(DSILoss).Most of the current methods are considering the problem from the perspective of input information,and how to find more separable clues is one of the main directions.Different from the traditional views,we intend to use the loss function to constrain the distribution of real face and spoofing face in the high-dimensional space,and improve the generalization of the model by controlling the distribution.At the same time,this paper put forward three principle of designing the loss function which included 1.The distribution of real face should be dense and the spoofing ones should be sparse;2.The union set of real face and attacks is the whole space;3.The center of two distributions should not be too far.Based on these principles and inspired by the loss function of face recognition,the DSI loss function which utilized limited datasets information for training is proposed to expand the domain of the spoofing face as much as possible and so as to enhance the generalization of the model,even the system meet the unknown types of attack can also accurately identify the spoofing faces.Several experiments show that the method is effective,and the visualization results further illustrate the rationality of the hypothesis.(3)Based on relevant research results,this paper also builds a Face AI library which provide a simple python API included face detection,face landmarks detection,face recognition,face anti-spoofing detection and so on to accelerate the progress of related research projects.