Research On Voiceprint Biometric Template Protection Methods

Biometric authentication technology can be used to authenticate the entity based onphysiological and behavioral characteristics measured automatically. It has become thegradual boost of the need of identity authentication in information society. With theapplication of biometric authentication system in a deep-going way, the importance ofbiometric transmission and storage security issues will be seen in the near future. Biometrictemplate protection technology provides a feasible solution for the security of biometricmodality itself by combining biometric measure with information security technology and hasarised worldwide research interests. Voiceprint template protection has important theoreticaland applied research value based on the advantages and the need of the voiceprint-basedauthentication. This thesis mainly focuses on the effective integration way between voiceprintand classic biometric template protection technology. Voiceprint template protection methodsalso are proposed based on information security technology. The main contributions of thisthesis are as follows:(1) A voiceprint-based authentication platform named SVoiceprintAuthen is developedwith MFCC features and three groups of models (VQ, GMM, GMM-UBM). Experiments oftext-independent speaker verification and voiceprint template protection can be bothimplemented on this platform.(2) For the leakage of original biometric by orthogonal transformation, the idea ofdimension reduction is put forward. A voiceprint template protection method based onrandom projection is proposed with the defined formalized representation of randomprojection. In the enrollment stage, voiceprint data are projected onto a random space, aGaussian mixture model (GMM) is then constructed, and the corresponding model parametersare stored as a template. In the verification stage, voiceprint data to be verified are matchedwith the model base data in the same random space. Moreover, the performance and securityof the proposed method are theoretically analyzed. The minimum dimension of the randommatrix is given according to Johnson-Lindenstrauss lemma and the maximum dimension ofthe random matrix for security protection performance is given according to l-securitylemma. Experimental results demonstrate the validity of the proposal for VQ, GMM,GMM-UBM and the analytical results of dimension reduction. With the same dimension ofrandom matrix, the performance of GMM is approximately maintained, while VQ may resultin a performance degradation of more than8%, which means that the random projection ismore suitable for the template protection of GMM voiceprint-based authentication systems. (3) A voiceprint template protection method based on feature disturbance and dynamicprojection is proposed for the disadvantages of the algorithm in part (2) which can’t keep theperformance and can’t enhance discrimination between the templates. In the enrollment stage,voiceprint data after feature disturbance are trained for the codebook. Then the codewords aredynamically mapped to different random spaces decided by the every element in thecodeword. In the verification stage, voiceprint data to be verified are disturbanced and thenmapped to the random spaces of the codewords to calculate Euclidean distances. Featuredisturbance can be used to enhance discrimination between the templates so as to improve thecancelability; while dynamic projection can be used to increase the uncertainty of thetransformation and strengthen the binding relationship between the template and the key so asto improve security. A set of measures for security evaluation of template transformationmethod are defined according to typical attacks and a detailed analysis of the security is thenconducted. Experimental results demonstrate that ROC curve is approximately diagonal whendifferent random matrixes are used to respectively map the codewords and the voiceprint,which shows the cancelability of the proposal.(4) A voiceprint template protection method based on fuzzy vault and random projectionis proposed for the disadvantages with no security evaluation measures of random projection.In the enrollment stage, voiceprint data after random projection are trained for the codebook.Then the chaff-points are added into the codewords to construct a voiceprint fuzzy vault. Inthe verification stage, voiceprint data to be verified after random projection are matched withthe voiceprint fuzzy vault in the same random space. The Euclidean distances between vectorsare computed and then minimum distortion measurements of the genuine codewords arefiltered by the binary indexes. The average minimum quantization error is calculated forauthentication. The advantages of the proposal are analyzed in detail such as overcoming thecross-matching and free-space-constrained of fuzzy vault, which improve the cancelabilityand security. Binary bits calculation instead of the prime comparision to filter minimumdistortion measurements can improve the authentication efficiency. The experimental resultsalso show the validity of the proposal for VQ and GMM.(5) On the vector element level, a voiceprint template protection method based on realhomomorphic encryption is proposed with the advantages of ciphertext calculation supported.In the enrollment stage, codebook elements are encrypted by homomorphic encryptionalgorithm in the range of the real. In the verification stage, components’ absolute values ofEuclidean distance between encrypted voiceprint vectors and codebook are calculated in theciphertext domain. Decision is then made according to the average minimum quantization error which is calculated by decrypting homomorphically component addition results ofEuclidean distance. Analysis demonstrates that the proposal can resist ciphertext-only attacks,but can’t resist known-plaintext attacks. The experimental results show the proposal can keepthe same performance of authentication system as in the original feature domain.(6) A voiceprint template protection method based on improved homomorphicencryption is proposed for the disadvantages of the algorithm in part (5). An improvedhomomorphic encryption algorithm in the range of the integer is constructed firstly,encrypting plaintext to non-negative integers. Analysis demonstrates that the ciphertext ofalgorithm do not leak the decimal fraction and sign information of plain text and can resistknown-plaintext attacks. Voiceprint and codebook elements are transformed into integersfirstly and then components of Euclidean distance are calculated in the ciphertext domain byimproved homomorphic encryption algorithm in the range of the integer. Decision is madeaccording to the average minimum quantization error which is calculated by decryptinghomomorphically component addition results of Euclidean distance. Authentication isconducted in the ciphertext domain so as to improve the security. The experimental resultsdemonstrate that integer transformation will probably reduce the performance of VQ andGMM. The results of GMM is worse than VQ’s. For the high performance requirements, alarge value of b can be taken in order to maintain authentication performance.