Design, implementation and evaluation of an EEG-based user authentication methodology using deep learning methods
Loading...
Date
2026-07-06
Authors
Ασπιώτη, Μαρία
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Biometric authentication has become an integral part of most people's daily lives, whether through unlocking a mobile phone using facial recognition or completing a purchase with a fingerprint. However, these traditional authentication methods are vulnerable to theft by third parties and offer no possibility of revocation if they are compromised. Electroencephalography (EEG) has emerged in the field of research as a novel authentication method, offering a unique, universal, revocable and inherently liveness-detecting biometric trait.
This thesis presents the design, implementation, and evaluation of a subject-specific EEG-based authentication methodology using deep learning techniques. A well-known dataset containing motor execution and motor imagery data was utilized for this implementation. Given that raw EEG signals are inherently noisy and prone to artifacts, a comprehensive preprocessing pipeline was developed.
A key contribution of this work is the successful application of Deep ConvNet, an architecture originally designed to decode EEG signals of imagined or executed movements, as a binary user authentication architecture. Furthermore, this thesis investigates the impact of data augmentation techniques in addressing limited data volume and class imbalance. Experimental results demonstrate that unaugmented data are insufficient for creating a secure system. The injection of Gaussian noise leads to models that frequently reject genuine users alongside impostors. In contrast, Sliding Window Augmentation provides an optimal balance of system security and user convenience, with both the False Acceptance Rate (FAR) and the False Rejection Rate (FRR) being close to zero.
Furthermore, the study compares the system's authentication capabilities based on the specific task being executed, specifically comparing motor imagery (MI) with motor execution (ME). Evaluations employing the Sliding Window Augmentation method, previously identified as the optimal data augmentation strategy, reveal that motor imagery (specifically, the imagined opening and closing of the left or right fist) outperforms actual execution. This MI task maintains system security (FAR < 2%) while reducing the FRR to approximately 3%.
In conclusion, this study finds that biometric systems utilizing EEG data, when combined with appropriate preprocessing, data augmentation, and deep learning models, can offer a secure and effective authentication methodology.
Description
Keywords
Electroencephalography (EEG), Biometric authentication, Brain-computer interface, Deep learning, Motor imagery, Motor execution, Signal processing, Data augmentation, Class imbalance, Deep convnet