A new training approach for deep learning in EEG biometrics using triplet loss and EMG-driven additive data augmentation

Abstract

One of the major challenges facing Electroencephalogram (EEG) in biometric systems is the high time-variability of the brainwaves, especially across different sessions. Electrical muscle activity, or Electromyogram (EMG), is considered one of the main noise sources that affect EEG repeatability due to the widely overlapping spectra. In this paper, we introduce a new training approach for deep learning to learn time-permanent and subject-unique embeddings towards an EEG biometric system. The proposed neural network is trained to minimize the triplet loss from EEG frames that are boosted with EMG-driven additive data augmentation. The EMG-like noise is synthetically generated and added to the original EEG batches during training, i.e. online augmentation, to reduce the memory usage specifically in large-scale EEG datasets. This framework is evaluated on a multi-session database that was collected specifically for EEG biometric evaluation under auditory stimulation and relaxation protocols from over 50 users. Adopting this approach showed significantly improved performance over brain-computer interface techniques and conventional deep learning schemes that use cross-entropy loss where a 6–20% improvement in the Correct Recognition Rate (CRR) and 3–7% decrease in the Equal Error Rate (EER) were achieved under cross-session setup. Besides, replicating session nuisance with the proposed framework demonstrated enhanced session-invariant EEG embeddings compared to adversarial learning methods with an average of 7.5% CRR improvements. Finally, our proposed training procedure attained a consistent performance on a longitudinal assessment setup where the time gap between enrollment and testing is almost 1 year. Under this setup, a CRR up to 99.8% and EER as low as 1.24% were achieved over a subset of 12 subjects. The achieved results demonstrate the effectiveness of the proposed training framework in improving the time-permanence of the EEG biometric traits under different protocols and testing scenarios.