Discrimination of Real and Deep Fake Videos using EEG Signals.

Area of Science:

• Neuroscience
• Computer Science
• Digital Forensics

Background:

• Deepfake technology poses significant ethical challenges and misinformation risks.
• Accurate deepfake detection is crucial but technically demanding.
• Existing detection methods often rely on visual artifacts, which deepfakes can circumvent.

Purpose of the Study:

• To investigate the efficacy of electroencephalography (EEG)-based approaches for deepfake detection.
• To analyze brainwave responses to real versus deepfake videos.
• To develop and evaluate machine learning models for classifying video authenticity using EEG data.

Main Methods:

• Recorded 64-channel EEG signals from 10 participants viewing 100 videos (50 real, 50 deepfakes).
• Preprocessed EEG data, removed artifacts, and applied Pearson's correlation.
• Extracted features using Wavelet Packet Decomposition (WPD) and Fast Fourier Transform (FFT).
• Trained five machine learning classifiers (e.g., SVM, k-NN) on extracted features.

Main Results:

• The Wavelet Packet Decomposition (WPD) method achieved a maximum accuracy of 94.16%.
• The Fast Fourier Transform (FFT) method, combined with a k-nearest neighbors model, reached 98.25% accuracy.
• Machine learning models effectively classified real and deepfake videos based on EEG features.

Conclusions:

• EEG-based analysis offers a potential passive method for deepfake detection.
• The high accuracy achieved demonstrates the viability of neurophysiological signals for identifying fabricated media.
• Further research is warranted to refine these EEG-based deepfake detection techniques.