Application Of Machine Learning To Face Anti-spoofing Detection

Biometric verification,as an important identification verification technology,is more and more widely used in the field of privacy and security.The use of human face biometrics for identity recognition has received extensive attention in academia and industry.However,because the traditional face recognition technology cannot effectively distinguish between genuine and fake faces,the problem of image spoofing brings many security risks.In order to overcome the problem of image recognition of authentic faces,this paper proposes three research methods of human face detection technology.The first method is based on traditional feature extraction and machine learning,fusing static Gabor wavelet features and dynamic LBP features,and using support vector machines(SVM)to classify them.However,traditional methods can only extract shallow features and cannot express deep semantic information.In order to solve this problem,this paper then proposes a second face detection method based on generative adversarial networks(GAN).The method is based on the deep convolution to generate the structural model of the adversarial network(DCGAN),and uses unsupervised learning for sample expansion(ie,data enhancement).The proposed model is trained to generate confrontation samples that are more difficult for classification tasks,so as to promote network learning and improve classification accuracy.This method can be used for any GAN model.By adjusting the output value of the discriminant model during the training process,a sample between two types(ie,an adversarial sample)is generated,which is used to expand the training set.SVM is used for classification.The experiment verifies that the data enhancement improves the accuracy.At the same time,it tests the effectiveness of the discriminative model's convolutional layer features for classification.Both of the above methods use supervised learning,but the training of supervised learning requires a large number of labeled data samples,and label calibration usually consumes a lot of manpower and material resources.Not only that,for some real and fake face samples that are difficult to distinguish,Label information is difficult to determine.In view of the above problems,the third method proposes a face live detection technology based on image repair model for semi-supervised learning.During the training process of image repair,it continuously classifies and learns from very few labeled samples.The image repair models used include generative models and discriminant models,among which discriminant models include local discriminant models and global discriminant models.Specifically,the generation model is mainly used to repair the missing image,the local discriminant model is used to discriminate the local consistency of the mask image,and the global discriminator is used to discriminate the global consistency of the complete picture and classify the labeled samples.The whole model can not only repair the picture,but also realize the semi-supervised learning task.The experiment proves the effectiveness and accuracy of semi-supervised learning based on image restoration for human face detection tasks by comparing existing algorithms.