【作者】 李鹏宇；

【导师】 庄伯金；

【作者基本信息】 北京邮电大学 ， 电子与通信工程， 2015， 硕士

【摘要】 如何安全可靠地对人的身份进行识别认证,从古至今都是一个十分重要的问题。古代社会通过印信、符节等方式来传递和认证身份信息,到近代社会以来,进行身份认证的方法更是多种多样。在众多身份认证的方式中,基于生物特征信息的认证方式由于其独特的安全性和便携性成为热点的研究领域之一。基于指纹、步态、声音、虹膜等生物信息进行个人身份认证的应用越来越多。但由于这些特征的固有特性,它们在防伪造、是否活体采集等方面存在不足,在一些特定的高危场景中使用还有缺陷。人的脑电信号除了具备成为生物特征的基本条件外,还具有隐蔽性、不可窃取性、不可仿制性、不可胁迫性以及必须活体等诸多独特的优势,更适合一些特定场合身份识别的应用。本文的主要工作集中在对现有的基于脑电信号的身份识别系统进行调研和改进,使其不仅能够进行集内样本的识别(闭集识别问题),还能拒识属于集外类别的样本(拒识问题),我们把这种既能做到集内识别,又能做到集外拒识的系统称为开集识别系统。与此同时,本文还在尽可能保证准确率性能的条件下,减少输入信号的时间长度。论文的具体工作包括以下几个部分：1.对自回归模型系数进行改进,并且提出了将功率谱密度、改进自回归模型系数和小波分解特征结合起来,通过线性判决降维后得到的高级特征作为开集识别系统的特征,取得了较好的效果；2.提出了采用度量学习的方法来处理高级特征,从而能够在系统准确率几乎不变的条件下,大幅度的降低采集数据的时间；3.提出一种两级分类器的方法,来构成分类系统,既能进行非法用户拒识,又能准确识别合法用户身份,提高了整个系统的性能。我们整个基于单通道脑电信号的身份开集识别系统,达到了错检率(FP)是6%的条件下,漏检率(FN)为5%,综合准确率为88.7%的性能效果。更多还原

【Abstract】 Authorizing personal identity in a safe way is important in the social life nowadays. In recent years, there are lots of methods to authorize identity. Among them, the method based on biometric information has become one of the most popular ways because of its security and convenience. There have been more and more researches in personal authentication based on fingerprint, iris, voice and gait. Compared with face, fingerprint, voice, iris, signature or other widely used biometrics, EEG has two distinct advantages. First, the active Electroencephalogram (EEG) must come from a living individual with a normal mental state, and an aggressor cannot force the person to provide the ideal EEG signals as those recorded in normal states. Second, EEG is hard to mimic. Therefore, EEG can be an excellent complement modality to the existing biometric systems which are prone to forgery.The paper focuses on the research in personal identification with the single channel EEG signal. An improved algorithm is propose that can not only find who he is if the subject belongs to the data set (closed-set identification) but also can reject the subject if he/she is out of the data set (rejection). Besides, metric learning is used in our algorithm to reduce the required EEG signal’s recording duration, and keep almost the same auucracy.tThe main contributions are as follows:1. The auto-regression (AR) model is improved, and feature level fusion is proposed including the power spectral density, AR model, and wavelet feature. Furthermore, the linear discriminant analysis (LDA) is used to reduce the dimensions of the cascaded feature. Experimental results shows the effectiveness of the proposed method.2. Metric learning is introduced to reduce the required EEG signal’s recording duration, and keep almost the same accuracy.3. A two-level classifier with naive Bayes classifier and K nearest neighborhood (KNN) is proposed, which can not only identify the legal person in the data set, but also reject the illegal person out of the data set. Finally, we achieved that the false positive rate is6%while the false negative rate is5%, and the final open=set identification accuracy rate is88.7%.更多还原