Detecting Deepfakes with Super-Resolution EEG.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Electroencephalogram (EEG) signals are vital biomarkers for brain activity.
• Spatial resolution, determined by channel count, is crucial for understanding brain functions.
• Low-resolution EEG limits detailed analysis of neural processes.

Purpose of the Study:

• To develop a deep autoencoder for generating super-resolution (SR) EEG signals from low-resolution (LR) data.
• To evaluate the proposed method's performance against traditional interpolation techniques.
• To assess the utility of SR EEG in deepfake classification tasks.

Main Methods:

• A deep autoencoder architecture was designed to process 32-channel LR EEG signals.
• The model was trained to output a 63-channel SR EEG signal.
• Performance was quantified using Mean Squared Error (MSE), correlation, and Peak Signal-to-Noise Ratio (PSNR).
• LR, SR, and original high-resolution (HR) EEG data were used for deepfake classification via Naive Bayes.

Main Results:

• The autoencoder approach significantly outperformed bilinear interpolation, reducing MSE by 74.69%.
• Correlation increased by 27.70% and PSNR by 25.19% with the proposed SR method.
• Deepfake classification accuracy using SR EEG (61.13%) closely matched HR EEG (62.21%), surpassing LR EEG (58.35%).

Conclusions:

• The deep autoencoder effectively generates high-resolution EEG signals from low-resolution inputs.
• The SR EEG approach demonstrates potential for enhancing brain activity analysis and applications like deepfake detection.
• This method offers a viable solution for improving spatial detail in EEG data acquisition.