Data-Efficient EEG-Based Person Identification via Time-Interval-Based Multi-Positive Contrastive Learning.

Electroencephalography (EEG) is an emerging biometric trait because it is difficult to steal or forge and requires acquisition from living subjects. However, many EEG-based person identification methods rely on large enrollment datasets, increasing acquisition burden and limiting practical deployment. To address this issue, we propose a data-efficient person identification framework based on time-interval-based multi-positive contrastive learning (TIB-MPCL). For each one-second EEG segment, overlapping, adjacent, and non-adjacent positive samples are constructed to exploit temporal diversity. These samples are jointly optimized through contrastive learning, together with an alignment module that promotes intra-subject compactness and inter-subject separability. Experiments on the 109-subject PhysioNet dataset demonstrate strong performance under both sufficient- and limited-sample conditions. TIB-MPCL achieves an average accuracy of 99.05% using all available training samples and 83.86% using only 10 one-second samples per subject, with accuracies above 91% for motor execution and imagery. Cross-session experiments on SEED-IV further evaluate robustness to session-dependent variability, achieving an average accuracy of 83.79% when the test session is completely excluded from training.