Stable EEG Biometrics Using Convolutional Neural Networks and Functional Connectivity

Despite some positive results, research on electroencephalography (EEG) biometrics is still in an exploratory phase. The sensitivity of EEG to physical and mental states leads to considerable intra-individual variations that can have a serious negative impact on biometric performance. One open question is the stability of EEG biometrics in relation to the consistency of performance across diverse human states. This paper investigates the idea of using functional connectivity (FC) and convolutional neural networks (CNN) for stable EEG biometrics. Specifically, the proposed learning model consists of two modules: an FC module which is designed to capture dynamic coupling relationships between brain regions, providing a bivariate measurement that is more stable than current methods relying on univariate signals or features extracted from single channels; and a CNN module to automatically learn inherent FC representations that exhibit unique patterns to each individual. The fusion of FC and CNN provides effective biometric identifiers and strong discriminatory power that relaxes the reliance of current methods on human states or sensory stimulation. Evaluation of the model uses ongoing EEG in different human states and both the identification and authentication scenarios. Results validate the effectiveness of the model in learning identity-bearing information from EEG in both biometric scenarios. Compared to current methods, the improvement brought by the proposed learning model is clear, especially in handling EEG signals in diverse states.