Adversarial Deep Learning in EEG Biometrics.

Area of Science:

• Neuroscience
• Machine Learning
• Biometrics

Background:

• Deep learning for electroencephalographic (EEG) person identification struggles with temporal correlations and session-specific variability.
• Existing methods often lack robustness due to training and evaluation on single-session EEG data, limiting longitudinal usability.

Purpose of the Study:

• To develop a deep learning approach for robust person identification from EEG data that overcomes session-specific variability.
• To learn session-invariant, person-discriminative representations from EEG signals for improved longitudinal accuracy.

Main Methods:

• An adversarial inference approach was integrated into a deep convolutional neural network architecture.
• The model was trained to learn representations invariant to recording session variations while maintaining person-specific discriminative power.
• The method was evaluated using longitudinally collected EEG data, analyzing half-second EEG epochs.

Main Results:

• The proposed adversarial inference method demonstrated improved performance in person identification compared to baseline methods.
• The learned representations exhibited enhanced robustness against variations across different EEG recording sessions.
• Empirical assessments confirmed the effectiveness of the approach for longitudinal EEG-based person identification.

Conclusions:

• Adversarial inference offers a promising strategy for learning session-invariant representations in EEG-based biometrics.
• The developed method enhances the longitudinal usability of deep learning models for person identification from brain activity.
• This work contributes to more reliable and robust brain-computer interfaces and security applications.