Font Size: a A A

Research On Perceptual Information Based Multi-modal Biometric Fusion Techniques

Posted on:2010-12-09Degree:DoctorType:Dissertation
Country:ChinaCandidate:Z F WangFull Text:PDF
GTID:1118360302965522Subject:Information security
Abstract/Summary:PDF Full Text Request
Biometric recognition technique has been widely used in information security andother fields because it can balances the con?ict between the system security and the usersexperience to some extent. Due to the inherent feature of single-modal biometric method,it is facing some problems in security and performance, which can be overcome by multi-modal biometrics for the better practicability, higher security and better performance. Thedevelopment of multi-modal biometrics would enrich the biometric recognition field andprovide more reliable and secure identity authentication scheme for information securityapplications. An effective way to implement multi-modal biometrics is to perform fusionon the feature-level which derives the most discriminative information from the originalmultiple feature sets and eliminates the redundancy among different feature sets. How-ever, current research in feature-level multi-modal biometrics is quite arbitrarily. Thereason lies in two aspects: the features from different modal are various in the data for-mat and data distribution; there is no thorough analysis on these features and no guidingframework.Referring to the information processing mechanism of human vision perceptual sys-tem, this dissertation classifies biometric features into two levels: the sensory featuresand the perceptual features based on the analysis of the features of multi-modal samples.Then an extended feature-level multi-modal fusion model is proposed, which is a reason-able guiding framework for multi-modal biometrics. Under the guiding framework, thesensory feature level fusion, perceptual feature level fusion and the crossing level fusionare investigated through the experiments on the fingerprint, iris and face. Finally the se-curity issues in multi-modal biometric system are discussed, an evaluation analysis forsample partial leakage is presented and a template protection algorithm based on adaptivenon-uniform quantization is brought out.The main contributions of the dissertation are described as follows:(1) An extended feature-level multi-modal fusion model is firstly proposed. Thefiner feature-level fusion framework is brought out based on the investigation on the sen-sory features and perceptual features. The feature-level fusion issue is deducted as theissue of sensory feature fusion, the perceptual feature and the crossing feature fusion, serving as a theoretical guideline for the feature selection and fusion strategy selection.(2) Aiming at the sensory feature extraction of typical biometric modals, the pre-process methods and sensory feature extraction algorithms of fingerprint, iris and faceare investigated. Based on the extracted sensory features, a weighting rule based fusionmethod is involved to implement the fusion between fingerprint-iris, fingerprint-face andface-iris. The experimental results demonstrate the feasibility of the sensory feature ex-traction methods.(3) Considering the perceptual features as the point set in the feature subspace, thefeature subspaces of perceptual features from different modals are normalized by z-scoremodel based on the analysis on the properties of perceptual features space. Then the ex-tended common vector algorithm in complex field is proposed for multi-modal perceptualbiometric feature fusion. Two methods of solving common vector are presented and theirequivalence is proved. Finally the fusion method based on complex field common vectoris performed on fingerprint, iris and face. Experimental results show the validity of thefusion method.(4) To solving the nonlinear problem in the fusion of sensory feature and percep-tual feature, two kernel based extended fusion algorithms are proposed: extended kernelprinciple component analysis algorithm (EKPCA) and extended kernel fisher discrimi-nate analysis algorithm (EKFDA). The theoretical principle of EKPCA in both central-ized and non-centralized sample sets are derived. Then the cross-level fusion algorithmsfor the three typical modals are presented. Experimental results show that the proposedalgorithms outperform the conventional multi-modal biometric algorithms greatly.(5) Finally, the security issue in multi-modal biometric systems is discussed. Thesecurity threat parameter and discriminability parameter are defined, and used to evaluatethe impact of the partial leakage of biometric samples on the security performance ofthe system. The security performances of the fusion algorithms proposed in this thesisand a few representative traditional fusion algorithms are evaluated under the conditionof partial leakage. The experimental results demonstrate the EKFDA proposed in thisthesis exceeds the others. Then an adaptive non-uniform quantization based multi-modalbiometrics template protection scheme is proposed. By defuzzying the fusion features,the template is stored and matched based on the cryptographic hashing. The experimentalresults demonstrate the template protection scheme can increases the security of the multi- modal biometric system without debasing the performance of the multi-modal biometricsystem.
Keywords/Search Tags:Multi-modal biometrics, Sensory-perceptual feature, Extended common vector, Extended KPCA, Extended KFDA, Template protection
PDF Full Text Request
Related items