Research On The Key Technologies Of Secure Identity Verification Based On Face Biometrics

Face verification is one of typical biometric recognition technologies,with the advantage of non-contact and convenient capturing.It plays an important role in many application scenarios.In recent years,with the development of E-commerce and so on,biometrics such as human face are bound with more and more personal property and private information.Human face is non-renewable;what's more,it is easy to obtain users' face information from the social networks.As a result,there are security risks in face recognition system.It is important to study and enhance the security of face recognition system.It is also the precondition for wide application of face recognition system.This thesis studies on the following security problems in the face recognition systems: 1.Malicious attackers use the fake faces of legitimate users such as photos or videos to intrude their personal accounts.Hence it's necessary to design the effective face anti-spoofing methods.2.Hackers steal the face features that be saved in the database to attack the system or to use them for other purposes.Therefore,face biometrics should be protected.In order to deal with the security risks mentioned above,a series of creative algorithms have been proposed in this thesis.The main works can be summarized as follows:A multi-perspective dynamic features based face anti-spoofing method is proposed.Existing face anti-spoofing methods are affected by different cameras and display devices,and their performances degraded in cross-database test.Focusing on this problem,we propose a face anti-spoofing scheme to fusing multi-perspective dynamic features.One feature is the globally extracted temporal motion pattern,which maps the local and global motion information of the face in the video into a single image by analyzing the variation of pixel value throughout the video.The motion patterns of genuine and fake faces are much different.Furthermore,such differences are independent of cameras and display devices.The other feature is the visual rhythm of noise patterns,which from single imaging differs significantly from that from secondary imaging.Therefore,visual rhythm technique can combine multi frames on a single image.These two types of features are fused at the different levels for classification.Cross-database tests are conducted using CASIA-FASD as the training set and Replay-Attack database as the testing set.The experimental results show that the proposed scheme outperforms state-of-the-art algorithms.An approach is proposed to generate adversarial examples for face antispoofing with the minimum perturbation dimensions and limitation on perturbation interval.Face anti-spoofing algorithm based on deep learning is vulnerable to adversarial examples.It means that,a small and imperceptible perturbation on input image can confuse the deep learning based face anti spoofing algorithm.Therefore,it is necessary to analyze the mechanism of generating the adversarial examples,so that the face-spoofing detection algorithms can be more robust.Motivated by the above,we propose an approach to generate the adversarial examples for face-spoofing detection by combining the minimum perturbation dimensions and visual concentration.In our approach,perturbation is concentrated on a few pixels in a single color component,and the adversarial examples are generated iteratively.What's more,the intervals between pixels are constrained—according to the visual concentration.The generated adversarial examples can be perceived by human with low probability.The adversarial examples generated from the proposed approach can defraud the deep neural networks based classifier with only 1.36% changed pixels on average.Furthermore,human vision perception rate of the proposed approach decreases about 20% compared with the existing algorithm DeepFool.A multi-regional convolutional neural networks is proposed for face antispoofing.The existed convolutional neural networks classifiers over-emphasize local area,and they cannot utilize the information in the whole face image effectively.Based on the consistency of each local patch,we propose a novel face anti-spoofing scheme based on multi-regional convolutional neural network.The convolutional layer instead of the fully connected layer is utilized.And not the single label but the feature maps is output.What's more,the concept of local classification loss is proposed to local patches for training.The experimental results show that,the proposed method is effective to both traditional attacks and adversarial examples attack.A secure face verification scheme based on homomorphic encryption and deep neural network is proposed.To ensure both the accuracy and safety of face templates in face-based verification systems,we propose an encrypted face verification system.In this system,face features are extracted based on deep neural network,and then encrypted with the Paillier algorithm.The framework of the system involves three parties: the client,data server,and verification server.The data server saves the encrypted user features and user ID,the verification server performs computation,and the client is responsible for collecting a requester's information and sending it to the servers.The information is transmitted among parties as cipher text,which means that no party knows the private keys except for the verification server.It makes sure the safety of data in data server.The proposed schemes with two deep convolutional neural network structures are tested on LFW and faces94 dataset.The extensive experimental results show that the proposed approach can improve the safety of the system with little decrease in accuracy.What's more,a simple replacement of key can make facial feature revocable.Therefore,the proposed system is effective with respect to both the security and high verification accuracy.In general,we study the safety problems of face-based authentication system from different viewpoints.And we propose some creative algorithms,which achieve performances better than state-of-the-art on multi public databases.