Detecting Deepfakes with Super-Resolution EEG

Abstract

Electroencephalogram (EEG) signals play a crucial role as biomarkers of brain activity, providing valuable insights into neural processes. The spatial resolution of EEG, determined by the number of channels, is essential for obtaining a comprehensive understanding of brain functions. However, low-resolution EEG systems pose a significant challenge in capturing detailed brain processes. In this paper we propose a deep autoencoder structure that essentially takes in the information from low resolution (LR) 32 channels of EEG and outputs a super-resolution (SR) 63 channel EEG signal. Our proposed approach, draws its inspiration from image super-resolution literature, and greatly improves on traditional interpolation approaches such as bilinear interpolation, with a 74.69% reduction in mean-squared error (MSE), 27.70% increase in correlation, and 25.19% increase in peak signal-to-noise (PSNR) ratio. Furthermore, the resulting LR, SR, and original high resolution (HR) signals were used within a Naive Bayes for deepfake classification; this yielded similar SR and HR classification results (61.13% and 62.21%, respectively), whereas the LR classification lagged behind both, with an average accuracy of 58.35%. This serves as a proof-of-concept for generating higher resolution EEG from low resolution data to channels as high as 63 with a simple autoencoder based structure, with application to deepfake detection.