EEG-Based Person Authentication Modelling Using Incremental Fuzzy-Rough Nearest Neighbour Technique
High level security has nurtured the arisen of Electroencephalograms (EEG) signals as a noteworthy biometrics modality for person authentication modelling. Modelling distinctive characteristics among individuals, especially in a dynamic environment involves incremental knowledge updates from time t...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English English |
| Published: |
2016
|
| Subjects: | |
| Online Access: | http://eprints.utem.edu.my/id/eprint/18351/1/EEG-Based%20Person%20Authentication%20Modelling%20Using%20Incremental%20Fuzzy-Rough%20Nearest%20Neighbour%20Technique.pdf http://eprints.utem.edu.my/id/eprint/18351/2/EEG-Based%20Person%20Authentication%20Modelling%20Using%20Incremental%20Fuzzy-Rough%20Nearest%20Neighbour%20Technique.pdf http://eprints.utem.edu.my/id/eprint/18351/ https://plh.utem.edu.my/cgi-bin/koha/opac-detail.pl?biblionumber=100138 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | High level security has nurtured the arisen of Electroencephalograms (EEG) signals as a noteworthy biometrics modality for person authentication modelling. Modelling distinctive characteristics among individuals, especially in a dynamic environment involves incremental
knowledge updates from time to time. K-Nearest Neighbour (KNN) is a well-known incremental learning method which applies First-In-First-Out (FIFO) knowledge update strategy. However, it is not suitable for person authentication modelling because it cannot preserve the representative EEG signals patterns when individual characteristics changes over time. Fuzzy-Rough Nearest Neighbours (FRNN) technique is an outstanding technique
to model uncertainty under an imperfect data condition. The current implementation of FRNN technique is not designed for incremental learning problem because there is no update
function to incrementally reshape and reform the existing knowledge granules. Thus, this research aims to design an Incremental FRNN (IncFRNN) technique for person
authentication modelling using feature extracted EEG signals from VEP electrodes. The IncFRNN algorithm updates the training set by employing a heuristic update method to
maintain representative objects and eliminate rarely used objects. The IncFRNN algorithm is able to control the size of training pool using predefined window size threshold. EEG signals such as visual evoked potential (VEP) is unique but highly uncertain and difficult to process.There exists no consistant agreement on suitable feature extraction methods and VEP electrodes in the past literature. The experimental comparison in this research has suggested
eight significant electrodes set located at the occipital area. Similarly, six feature extraction methods, i.e. Wavelet Packet Decomposition (WPD), mean of amplitude, coherence, crosscorrelation, hjorth parameter and mutual information were used construct the proposed person authentication model. The correlation-based feature selection (CFS) method was used to select representative WPD vector subset to eliminate redundancy before combining with other features. The electrodes, feature extraction, and feature selection analysis were tested using the benchmarking dataset from UCI repositories. The IncFRNN technique was evaluated using a collected EEG data from 37 subjects. The recorded datasets were designed in three different conditions of ambient noise influence to evaluate the performance of the proposed solution. The proposed IncFRNN technique was compared with its predecessor, the
FRNN and IBk technique. Accuracy and area under ROC curve (AUC) were used to measure the authentication performance. The IncFRNN technique has achieved promising results. The
results have been further validated and proven significant statistically using paired sample ttest and Wilcoxon sign-ranked test. The heuristic incremental update is able to preserve the core set of individual biometrics characteristics through representative EEG signals patterns
in person authentication modelling. Future work should focus on the noise management in data acquisition and modelling process to improve the robustness of the proposed person authentication model. |
|---|