Self-Organizing Clustering by Growing-SOM for EEG-based Biometrics

The use of electroencephalography (EEG) for bio-metric recognition, in particular for verification systems, has increased in the last years, due to some advantages that EEG signals present when used as signatures, as compared to other identifiers. In this paper we explore the use of unsupervised adaptive learning as a tool for enhancing the features representing each possible subject in a biometric system, in order to improve its performance. To do so, we designed three different frameworks based on Self Organizing Maps (SOM) neural networks, and compared their performance with a base model using no enhancement. Our experiments, using different input tasks and two combinations in the number of channels, with data obtained from two public EEG databases, showed that a SOM with Dynamic Structure (GSOM) obtained the best Equal Error Rate (EER). Such EER was 0.08 ± 0.04 when using as input the counting task of a public database provided by the University of Colorado, and an EER of 0.11 ± 0.04 was obtained for the rotation task in the same database. We also assessed our frameworks using the public database BIOMEXDB, provided by INAOE, where we also found that GSOM outperformed other state-of-the-art works.