Discrimination of Real and Deep Fake Videos using EEG Signals

Deepfake technology can create highly realistic fabricated videos, presenting serious ethical concerns and threats of misinformation. Reliably distinguishing deepfakes from genuine videos is therefore critical yet challenging. This study explored electroencephalography (EEG)-based deepfake detection by analyzing EEG responses in 10 participants viewing 100 videos (50 real, 50 deepfakes). Signals were recorded with a 64-channel system. Following standard preprocessing and artifact removal, data was analyzed using Pearson's correlation and features from the selected channels were extracted using wavelet packet decomposition (WPD) and fast Fourier transform (FFT). Five machine learning classifiers (support vector machine, k-nearest neighbors, etc.) were trained on these features to classify real versus deepfake videos. The WPD approach achieved a maximum accuracy of 94.16%, while the FFT method attained 98.25% accuracy using a k-nearest neighbors model. These EEG-based models demonstrate potential for passively detecting deepfakes, meriting further research.