基于随机映射的声纹模板保护方法  

Random Projection-Based Template Protection for Voiceprint-Biometric Systems

作　　者：朱华虹[1] 贺前华[1] 李艳雄[1] 张雪源[1] 

机构地区：[1]华南理工大学电子与信息学院,广东广州510640

出　　处：《华南理工大学学报（自然科学版）》2013年第5期48-54,共7页Journal of South China University of Technology(Natural Science Edition)

基　　金：国家自然科学基金资助项目(60972132;61101160);广东省自然科学基金团队项目(9351064101000003);广东省自然科学基金博士启动项目(10451064101004651);华南理工大学中央高校基本科研业务费专项资金资助项目(2011ZM0029)

摘　　要：针对生物特征模板涉及用户隐私而易受到各种攻击的问题,在定义随机映射形式化表示的基础上,结合主流文本无关说话人识别技术,提出了一种基于随机映射的声纹模板保护方法.在注册阶段,将声纹特征映射至随机空间再训练高斯混合模型(GMM),并存储模型参数作为模板;在认证阶段,待认证的声纹特征在相同的变换域与模型库进行匹配.文中还给出了该方法在认证性能保持和安全性方面的理论分析.实验结果表明,适当降维能在提高安全性的同时近似保持GMM的身份认证性能,而目前基于矢量量化的认证性能下降超过8%,说明随机映射更适用于基于GMM的声纹认证系统的模板保护.Biometric templates are vulnerable to various attacks because they involve user＇ s privacy. In this paper, by combining the defined formalized representation of random projection with the mainstream text-independent speaker recognition, a template protection method based on random projection is proposed for voiceprint-biometric systems. In the enrollment stage, voiceprint characteristic data are projected onto a random space, a Gaussian mixture model （GMM） is then constructed, and the corresponding model parameters are stored as a template. In the verification stage, the voiceprint characteristics to be verified are matched with the model base data in the same random space. Moreover, the performance and security of the proposed method are theoretically analyzed. Experimental results show that suitable dimensionality reduction helps to improve the system security and approximately maintain the performance of GMM, while the existing vector quantization method may result in a performance degradation of more than 8% , which means that the random projection is more suitable for the template protection of GMM voiceprint-biometric systems.

关 键 词：声纹 模板保护 高斯混合模型 随机映射 

分 类 号：TP309[自动化与计算机技术—计算机系统结构]