| Summary: | Studies have shown that the electroencephalogram (EEC) recordings have unique pattern for each individual and thus have potential for biometric applications. There are two major problems associated with EEC biometrics. One is that the large EEG features size and the relatively limited EEG data size make it difficult to train a robust model; the other is that the signals from EEG scalp may not be reliable in many situations. Thus in this thesis we proposed new methods for increasing the accuracy and robustness of
EEC-based authentication systems.
First, to address the concern of the high dimensionality of EEC features, we proposed a novel dimension reduction method of EEG features based on the Fast Johnson-Lindeustrauss Transform (FJLT). We showed that this method has potential of mapping EEG features from a high dimension space
to a lower one while keeping discrimination power between the features of subjects. The features we used are Multivariate Autoregressive (mAR)
coefficients. We tested this method on a motor task related EEG data set. Second, to increase the reliability of scalp EEC signals, we employed
an Independent Component Analysis (ICA)-based approach in our authentication procedure, with the assumption that EEG recordings are linear
combinations of the underlying brain source signals. We estimated the Independent Components (ICs) from several physical regions on the scalp and determine the Dominating Independent Components (DIC) in the corresponding regions. Then we extracted the Univariate Autoregressive (AR)coefficients from DICs as features. We tested our algorithm on two data sets, a motor task related EEC data set and an EEC data set of P300
potential. The proposed algorithm appeared to be promising, and when applied to EEG data collected from different days yields better performance than other methods. This relative consistence over time is essential in person authentication systems. === Applied Science, Faculty of === Electrical and Computer Engineering, Department of === Graduate
|
|---|
