Yap, Hui Yen (2025) Adaptive transfer learning and word stimulation for robust EEG-based authentication. Doctoral thesis, Universiti Teknikal Malaysia Melaka.
![[img]](http://eprints.utem.edu.my/style/images/fileicons/text.png) |
Text (24 Pages)
Adaptive transfer learning and word stimulation for robust EEG-based authentication (24 pages).pdf - Submitted Version Download (1MB) |
|
Text (Full Text)
Adaptive transfer learning and word stimulation for robust EEG-based authentication.pdf - Submitted Version Restricted to Registered users only Download (3MB) |
Abstract
Electroencephalogram (EEG)-based authentication has gained increasing attention as an alternative to conventional biometric systems due to its resistance to spoofing and privacy compliance. However, practical adoption remains limited, primarily due to high noise levels in consumer-grade EEG devices, high signal variation in different sessions, and the extensive training data requirements for deep learning models. Apart from ensuring biometric system performance, an EEG-based authentication system must also be user-friendly with a reasonable acquisition time to maintain user engagement. This study explores the feasibility of using consumer-grade EEG devices for authentication to address challenges such as noise and signal variability. It involves the design of a reasonably timed word-stimulation acquisition protocol to enhance signal reliability while minimizing cognitive fatigue. Additionally, due to the limited availability of training data, the performance of deep learning with transfer learning using pre-trained CNN models is investigated. The frequency spectra of the preprocessed EEG signals were extracted and used as input for pre-trained models. Experiments were conducted on a database of 30 subjects recorded over two separate sessions to evaluate the performance of the proposed method. Baseline evaluations compared pre-trained CNN models against traditional classifiers: SVM and k-NN. The results show that deep learning provides better performance within the same session. However, all methods, including pre-trained CNN models, SVM, and k-NN, experience performance degradation when tested on a different session dataset, revealing the challenge of EEG variability. In order to address this issue, an adaptive retraining strategy is proposed, which improves classification accuracy across sessions compared to direct deep learning transfer. These findings confirm the applicability of consumer-grade EEG devices for biometric authentication while addressing key challenges such as noise reduction, limited training data, and session variability. The proposed methodology contributes to the advancement of EEG-based biometric security, paving the way for practical deployment of EEG authentication systems in real-world applications.
| Item Type: | Thesis (Doctoral) |
|---|---|
| Uncontrolled Keywords: | EEG, Authentication, Transfer learning, CNN, Deep learning |
| Subjects: | Q Science Q Science > QA Mathematics |
| Divisions: | Faculty of Information and Communication Technology |
| Depositing User: | Norhairol Khalid |
| Date Deposited: | 21 Jan 2026 07:05 |
| Last Modified: | 21 Jan 2026 07:05 |
| URI: | http://eprints.utem.edu.my/id/eprint/29389 |
| Statistic Details: | View Download Statistic |
Actions (login required)
 |
View Item |