The Study Of Face Anti-spoofing Based On 3D Convolutional Neural Network

Spoofing face can be used to deceive face authentication systems for illegal purposes,and thus it presents threats to security.Therefore,it is great significance to investigate detection for spoofing face.Many existing literature focus on the study of photo attack.The research of video attacks was still insufficient.The 3D Convolutional Neural Network(CNN)has the characteristics of deep learning and can automatically learn the distributed feature representation of the image.Compared with 2D convolution,it can learn the motion information of continuous video frame.This paper through 3D convolution to realize video attack detection,which is based on the characteristics of 3D convolution neural network.Generally speaking,the main work is just as follow:(1)Through the deep study in neural network and living face detection,we learn the drawbacks of the existing general convolution neural network.It can only be two-dimensional image convolution.There are the following shortcomings of the existing artificial face detection algorithms.On the one hand,hand-crafted feature only has a good recognition rate for some kind of image,on the other hand,they require high professional knowledge.In this paper,3D convolution neural network is first proposed for living face detection based on the convolution of the video.(2)Aiming at the two open face anti-spoofing databases used in the experiment,a suitable 3D convolution neural network structure is designed,which includes the number of layers,the size and the number of convolution kernel.It is worth noting that t three alternative networks(Late,Slow,and Early)based on the order in which each network is sampled in the time dimension.Finally,through the experiment,we compare each network recognition rate and the number of network parameters to select the appropriate network.(3)In this paper,the multi-frame input is used for experimental comparison.The optimal input frame number is obtained by using the optimal network.In the experiments,we use two public databases,such as REPLAY-ATTACK and CASIA-FASD,to train the 3D convolutional neural network.The SVM linear classifier is trained by extracting the features of the final fully connected layer of the 3D convolutional neural network to classify the real face and the fake face.(4)Using the optimal network and the optimal frame number as input,the multi-scale internal database test and cross-database test of two public face anti-spoofing databases: REPLAY-ATTACK and CASIA-FASD are realized.The experiment is divided into five scales to complete the test,in the network input layer,the image is divided into 5 frames.After using the 3D convolution network to learn the characteristics of the image frame,the features of the last connection layer are extracted to train a SVM classifier to implement the classification of real face and fake face.Comparing with the texture feature analysis and 2D convolution method,the performance is greatly improved.And can be applied to living face detection of video attack.