Research On Individual Differences Of Multi-Paradigm Evoked EEG For Biometric Identification

In recent years, biometric recognition has received general concerns around the world, and become a frontal and hot topic in the information age. It plays an important role in many fields, including the maintenance of national security, personal information security, as well as aviation security, in addition to the area of military and medical. In order to meet some requirements for special application and compensate for the lack of the existing technology, the researchers are committed to developing new biometrics, in which EEG (Electroencephalogram) is a novel attempt. Researches on EEG-based biometric recognition are still at an initial stage in China and other countries and there are huge exploratory spaces in the aspect of design for evoked potentials paradigm, EEG feature extraction and pattern recognition algorithms.Based on EEG individual differences and taking into account the problem of oversimplified evoked potentials paradigm in the existing research, two kinds of tasks including multi-paradigms were designed and completed in this study: the first task included the resting, visual cognition, calculation, and movement imagination paradigms with twenty subjects; the second task was visual evoked P3 paradigm with eight subjects. In order to effectively extract the individual differences information in EEG, a series of feature extraction algorithms in time/frequency-domain was proposed according to the characteristics of each paradigm signal, including AR model, power spectrum in time-domain and frequency-domain, phase-locking value, and coherent average algorithms. Then, the above-mentioned features in each paradigm were classified by using support vector machine (SVM) to obtain the statistical classification accuracy of all samples. The results in this thesis showed that individual differences of evoked potentials were obviously more significant than that of resting EEG, especially those paradigms with more complicated tasks and higher participation level by subjects and closer correlation to the thinking task were proved to cause more obvious individual differences of EEG with up to 98%classification accuracy, which verified the feasibility of EEG to be used as a biometric.On this basis, in order to further improve system performance and identify efficiency, some preliminary exploration of feature selection and channel optimization was developed in this study, and three methods were proposed here, including genetic algorithm, Fisher discriminant ratio, and recursive feature elimination. Compared with the results before optimization, the recognition rate after optimization increased as well as the number of channels has obviously simplified, which may provide a novel idea for the individual difference analysis of EEG and for its practical design in the field of biometrics in the future.