Research On Face Recognition And Adversarial Examples In Deep Learning

Face recognition has become an important authentication technology due to its good non-invasiveness and natural existence characteristics,and it has been widely used.Since deep learning won the image recognition competition in 2012,using neural networks to solve face recognition problems can be more robust and get better accuracy than traditional algorithms,which is more and more popular amongthe researchers.However,the existing face recognition technology can still have a lot of room for improvement.This paper starts from the face recognition technology under deep learning,and explores the better detection and recognition algorithms based on the existing algorithms.In addition,the paper also focuses on the problem of adversarial examples attack in deep learning and researches on fake examples of faces.The following are the main research contents of this article.Aiming at the problem of missing detection and false detection of the original SSD algorithm on small sample face detection,consideri ng the idea of feature fusion,a feature-enhanced fusion SSD face detection algorithm based on the pyramid network(FPEF-SSD)is proposed.The algorithm based on the multi-scale feature layers extracted from the traditional SSD structure,and uses feature pyramids to fuse features of the upsampling layer and the size-invariant convolution layer.By fusing the abstracting high-level feature maps which are rich in semantic information and shallow high-resolution low feature maps which contain more details with feature cascade manner,the underlying feature map has richer semantic detail information,and the generated enhanced multiscale feature maps have better feature descriptions of small faces,moreover,the detailed information not easy to lose.Aiming at the problems of traditional triplet loss is difficult in selecting pairs and the large training set,combined with the idea of center loss,and introducing the discrete cosine transform method,a face recognition based on low-frequency discrete cosine transform and CT loss algorithm(LFDCTCT-FR)is proposed.The algorithm first improves the original triplet loss,and introduces central loss co-optimization,which can better reduce the intra-class spacing,and no longer need to carefully select the training pairs,which avoids the problem of convergence caused by the big data of the pairs.Based on the discrete cosine transform matrix,the low-frequency coefficients of the sample is fused with a single convolutional network feature by weight,which can better enrich the face features and enhance the discrimination between classes.Aiming at the problems of slow optimization of existing gradient-based image attack algorithms and unrobust attack effects,a Projected Gradient Descent of face adversarial examples attack algorithm based on hidden patches and momentum iterative is proposed,which is called HPMI-PGD.The algorithm introduces momentum method on the basis of the PGD attack method,which makes the iterative process converge faster and accelerates the distance from the original class.Besides,the generated interference can be presented in a more natural way through the constraints of the face accessory range,which is not easily to be suspected.