Characterization of the event-related potentials during GAN-based generation of EEG signals and their data augmented subject classification

One of the most crucial challenges in exploring deep neural networks for signal modeling and classification tasks is the availability of a large quantity of domain-specific data. The generative adversarial network (GAN) has emerged as an effective artificial data generation method. Motivated by that, we explore the generation of auditory event-related electroencephalographic (EEG) signals and their task-specific authentication in this work. For synthesizing subject-specific EEG signals, a conditional deep convolutional GAN is used. Apart from measuring the usual signal similarity, we also computed the correlation between the subject-wise event-related potentials (ERPs) corresponding to the real and synthetic EEG data. The characterization of the ERPs highlights that the GAN is not only able to learn the distribution of real EEG signals but also can preserve their temporal characteristics. Further, EEG biometrics experiments are also performed to verify the effectiveness of the synthesized EEG signals in data augmentation. It is noted that the averaged classification accuracy improves by augmenting the real data set with synthetic EEG signals.