Towards Robust Subject Identification From EEG Segments With Data Augmentation Techniques

Electroencephalography (EEG) has a wide gamut of information, from subject identity to their emotional states encoded in it. Given a segment of the EEG signal, our study aims to identify the subject in a robust way using the EEGChannel-Net architecture. To achieve this goal, we add gaussian noise to the given 32-channel EEG signal, and minimize the ℓ<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> distance between embeddings of the original signal and that of the noisy signal in addition to the categorical cross entropy loss for classification of the 26 subjects in the dataset. Addition of noise to each channel at a fixed signal-to-noise ratio (SNR) of 15 dB gave the best balanced accuracy of 98.27% in the validation set. Our study concludes by justifying the optimality of the 15 dB SNR and highlights the superiority of our proposed objective function.