Identity verification using voice and its use in a privacy preserving system

Official URL: http://192.168.1.20/record=b1226358 (Table of Contents)

Since security has been a growing concern in recent years, the field of biometrics has gained popularity and became an active research area. Beside new identity authentication and recognition methods, protection against theft of biometric data and potential privacy loss are current directions in biometric systems research. Biometric traits which are used for verification can be grouped into two: physical and behavioral traits. Physical traits such as fingerprints and iris patterns are characteristics that do not undergo major changes over time. On the other hand, behavioral traits such as voice, signature, and gait are more variable; they are therefore more suitable to lower security applications. Behavioral traits such as voice and signature also have the advantage of being able to generate numerous different biometric templates of the same modality (e.g. different pass-phrases or signatures), in order to provide cancelability of the biometric template and to prevent crossmatching of different databases. In this thesis, we present three new biometric verification systems based mainly on voice modality. First, we propose a text-dependent (TD) system where acoustic features are extracted from individual frames of the utterances, after they are aligned via phonetic HMMs. Data from 163 speakers from the TIDIGITS database are employed for this work and the best equal error rate (EER) is reported as 0.49% for 6-digit user passwords. Second, a text-independent (TI) speaker verification method is implemented inspired by the feature extraction method utilized for our text-dependent system. Our proposed TI system depends on creating speaker specific phoneme codebooks. Once phoneme codebooks are created on the enrollment stage using HMM alignment and segmentation to extract discriminative user information, test utterances are verified by calculating the total dissimilarity/distance to the claimed codebook. For benchmarking, a GMM-based TI system is implemented as a baseline. The results of the proposed TD system (0.22% EER for 7-digit passwords) is superior compared to the GMM-based system (0.31% EER for 7-digit sequences) whereas the proposed TI system yields worse results (5.79% EER for 7-digit sequences) using the data of 163 people from the TIDIGITS database . Finally, we introduce a new implementation of the multi-biometric template framework of Yanikoglu and Kholmatov [12], using fingerprint and voice modalities. In this framework, two biometric data are fused at the template level to create a multi-biometric template, in order to increase template security and privacy. The current work aims to also provide cancelability by exploiting the behavioral aspect of the voice modality.