EEG-Based Person Authentication using Variational Universal Background Model

Abstract

EEG-based person authentication is an important means for modern biometrics. However EEG signals are well-known for small signal-to-noise ratio and have many factors of variation. These variations are caused by intrinsic factors, e.g. mental activity, mood, and health conditions, as well as extrinsic factors, e.g. sensor errors, electrode displacements, and user movements. These create complex variations of source signals going from inside our brain to the recording devices. This thesis first aims at developing a variational inference framework to learn a simple latent representation for complex data. Next, a
variational universal background model is created to represent the latent space of all possible users. This representation is proven useful for score normalising and is shown to improve the performance of the authentication system. Extensive experiments show the advantages of our proposed framework.
The contributions of this thesis to EEG-based person authentication include:
1. Application of variational auto encoder to EEG-based person authentication
2. Developing a novel variational universal background model for learning user models
3. A detailed performance comparison between Gaussian, GMM, VAE, VGM and
their UBM versions

Date of Award	2019
Original language	English
Supervisor	Dat TRAN (Supervisor) & Wanli MA (Supervisor)