EEG-Based Personal Identification by Special Design Domain-Adaptive Autoencoder

Area of Science:

• Neuroscience
• Machine Learning
• Biometrics

Background:

• Individual brain activity patterns from electroencephalogram (EEG) offer a novel biometric source.
• Autoencoders automate feature extraction, a critical step in biometric identification.
• Domain-adaptive autoencoders (DAAEs) are explored for enhanced feature extraction.

Purpose of the Study:

• To investigate the use of DAAE-extracted features for personal identification using EEG data.
• To evaluate two domain adaptation approaches within DAAE frameworks.
• To assess model performance across different EEG recording types and subject numbers.

Main Methods:

• Extracted latent features using two DAAE domain adaptation approaches.
• Employed KNN, ANN, SVM, and RF classifiers for personal identification.
• Evaluated models on longitudinal EEG data (resting state, auditory, cognitive stimuli) for 2, 5, and 7 subjects.

Main Results:

• Highest identification accuracy of 100% achieved by SVM with uniform referential DAAE features in a 2-subject task.
• RF classifier attained 99.84% accuracy with softmin referential DAAE features in a 2-subject task.
• Classification accuracy decreased as the number of subjects increased, indicating classification difficulty.

Conclusions:

• DAAEs effectively extract features from EEG for robust personal identification.
• SVM and RF classifiers demonstrate high efficacy in EEG-based biometrics.
• The number of subjects significantly impacts the accuracy of EEG-based personal identification systems.