Comparing Features Extraction Methods for Person Authentication Using EEG Signals

This chapter presents a comparison and analysis of six feature extraction methods which were often cited in the literature, namely wavelet packet decomposition (WPD), Hjorth parameter, mean, coherence, cross-correlation and mutual information for the purpose of person authentication using EEG signals. The experimental dataset consists of a selection of 5 lateral and 5 midline EEG channels extracted from the raw data published in UCI repository. The experiments were designed to assess the capability of the feature extraction methods in authenticating different users. Besides, the correlation-based feature selection (CFS) method was also proposed to identify the significant feature subset and enhance the authentication performance of the features vector. The performance measurement was based on the accuracy and area under ROC curve (AUC) values using the fuzzy-rough nearest neighbour (FRNN) classifier proposed previously in our earlier work. The results show that all the six feature extraction methods are promising. However, WPD will induce large vector set when the selected EEG channels increases. Thus, the feature selection process is important to reduce the features set before combining the significant features with the other small feature vectors set.