Research On Compatibility Problem Between FHE Over Integers And Kernel Based Recognition Algorithms

Biometrics is a kind of personal identification and recognition technology which is not easy to be copied.However,in the process of identification,the personal information may be intercepted and modified by the third party or deceived by other side of communication,when it's transmitted through the public network channel.Therefore,in the process of identification,the personal information and server data need to protect.There are two main challenges in solving the privacy-preserving:one is the compatibility problem between the cryptographic protocols and the kernel based recognition algorithms.The second is the low recognition rate and the low recognition efficiency caused by the constraint of encryption algorithm.In summary,we take the recognition algorithm of integer homomorphic algorithm as the research topic and the research aspects are as follows:(1)In this thesis,a compatible face recognition system is proposed by analyzing and studying on the DGHV encryption scheme over the integers,and selecting the linear kernel and polynomial kernel of SVM.Reference the pipeline architecture,in the training phase,the client image extraction,face detection,feature extraction,model training and send the encrypted model parameters to the server,the server receives and stores the models;In the recognition phase,client image extraction,face detection,feature extraction and feature vector encryption,server complete the process of face recognition by using the encrypted feature vector and classification models with parameters.The system only needs a single round of communication;(2)To meet the demand for computing partial homomorphism and ensure the security of the data transmission,the DGHV homomorphic encryption scheme is introduced,and the single bit-DGHV scheme is improved to k-bit-DGHV scheme.Because the DGHV scheme due to the constraint of a kernel based classification algorithm is only suitable for selecting the linear kernel and polynomial kernel.To keep the recognition rate,feature extraction using Gabor wavelet and K-PCA which retain data nonlinear and higher-order statistics;(3)Using Java,OpenCV and LIBSVM library to achieve distributed face recognition system and taking the ORL face database as the experimental object.Experiments show that the proposed scheme for cumulative matching rate was 91.90%.During the experiments,the lowest recognition rate was 89.04%and the highest was 97.62%,and a maximum recognition time was about 1294.5 ms.In this thesis,the face recognition system has a certain practical value with its faster identification of the appropriate time and acceptable recognition rate under the restricted condition of ensuring the privacy of the client and server.