Electroencephalogram (EEG)-based biometric sensing provides a promising pathway for secure and user-specific human authentication, but practical deployment remains limited by cross-session non-stationarity and potentially optimistic evaluation protocols caused by window- or event-level leakage. This study developed a cross-session EEG biometric sensing framework using interpretable spectral features and target-session calibration, with trial-level non-overlapping data partitioning to prevent overlap between training, calibration, and blind-test sets. Multi-channel EEG events were segmented into non-overlapping 2 s windows and aggregated into event-level samples to ensure strict isolation among training, validation, calibration, and blind testing subsets. Power spectral density, differential entropy, and log-variance features were evaluated using support vector machine, linear discriminant analysis, logistic regression, random forest, and a compact convolutional neural network designed for EEG decoding (EEGNet). Both closed-set identification and biometric verification were assessed using Rank-N accuracy, cumulative match characteristic curves, macro-F1, equal error rate, area under the curve, and true acceptance rate (TAR) at fixed false acceptance rate (FAR) levels. In the in-house cross-day dataset, limited target-session calibration improved Rank-1 accuracy from 44.82% to 98.70% for support vector machine (SVM) and from 19.93% to 99.94% for linear discriminant analysis (LDA). SVM with differential entropy achieved an equal error rate (EER) of 1.04 ± 0.60%, area under the receiver operating characteristic curve (AUC) of 0.9985 ± 0.0008, and TAR@FAR = 0.1% of 98.44 ± 1.06%. External validation on the public multi-session motor imagery dataset confirmed the calibration benefit. These results demonstrate strong closed-set identification and score-level verification performance under a cohort-wide, calibration-assisted cross-session experimental protocol.
Objective: To compare electroencephalography (EEG)-based subject-identification and open-set rejection performance across experimental paradigms while examining model robustness and cross-session generalization. Methods: The M3CV database was analyzed across resting state, transient sensory stimulation, steady-state sensory stimulation, P300 Oddball, and motor execution. Closed-set identification was evaluated across feature representations, conventional machine-learning models, and raw-EEG deep-learning baselines. Open-set performance was further assessed using an identity-first leakage-free nested procedure with 60 enrolled identities, 15 development unknown identities, and 20 final-test unknown identities. Results: Under the P4 within-session protocol, transient sensory stimulation achieved the highest Rank-1 accuracy (98.10%). Differential entropy and shrinkage linear discriminant analysis achieved mean Rank-1 accuracies of 98.01% and 98.29%, respectively. Under leakage-free nested open-set evaluation, transient sensory stimulation achieved DIR@FPIR = 5%, DIR@FPIR = 1%, and AU-OSCR values of 91.75%, 77.32%, and 97.36%, respectively. Deep-learning analyses confirmed strong within-session identity discrimination but showed model-dependent open-set paradigm rankings. In contrast, strict cross-session transfer produced marked degradation across all model families, with Rank-1 accuracy falling to approximately 1.9–5.4% and verification AUC to approximately 0.53–0.56; unsupervised target normalization provided little recovery. Conclusions: EEG paradigms exhibited strong but largely session-dependent identity discriminability. Transient sensory stimulation provided the most favorable within-session open-set performance under the primary conventional framework, whereas cross-session variability dominated paradigm-dependent differences under zero-shot transfer.