Evaluation of entropy features and classifier performance in person authentication using resting-state EEG

Abstract

Introduction: Resting-state electroencephalogram (EEG) presents a promising biometric modality due to its inherent liveness detection and resistance to spoofing, addressing critical vulnerabilities in conventional systems. However, its deployment faces fundamental trade-offs among accuracy, robustness, and hardware efficiency, particularly concerning optimal electrode configuration, discriminative feature extraction, and classifier generalization. Methods: To address these challenges, this study systematically evaluates thirteen entropy measures—including spectral entropy (SpEn), refined composite multiscale entropy, fuzzy entropy, and sample entropy (SaEn) etc.—alongside six classifiers (Quadratic Discriminant Analysis (QDA), Random Forests and Support Vector Machines etc.) for person authentication. Using 32-channel EEG recordings from 26 healthy participants under rigorous leave-one-out cross-validation (LOOCV), we quantified the impact of electrode selection and feature-classifier pairing. Results: Key findings demonstrate: QDA classifier achieved peak performance of 96.8% accuracy using 30 electrodes. Critically, a streamlined 9-electrode portable configuration retained 96.1% accuracy, demonstrating robust performance with reduced hardware requirements. SpEn measure exhibited superior biometric discriminability compared with other entropy measures, exceeding SaEn by 13.8 percentage points. Conclusion: These results advance the design of portable EEG biometric devices while highlighting entropy features’ scalability.