Biometric recognition based on electroencephalography (EEG), which captures intrinsic neural dynamics via scalp-recorded electrical activity, has shown great promise. Most existing studies leverage deep learning to enhance recognition performance. However, they typically rely on large amounts of subject-specific data or require retraining to accommodate new users, thereby limiting scalability and generalization. To address these issues, we propose the Multi-Prototype Adaptive Network (MPAN), a dual-branch framework tailored for few-shot EEG-based biometric recognition, enabling rapid adaptation to new subjects. Specifically, a Dual-Path Pretraining Network (DPPN) is first employed to learn discriminative and robust EEG representations through supervised contrastive and mutual distillation losses. Building upon the pretrained global and local metric subnets, the subsequent MPAN integrates complementary metric information via a multi-prototype fusion module, yielding reliable decision-making under few-shot conditions. Evaluated on two public EEG benchmarks, MPAN consistently surpasses state-of-the-art baselines, achieving identification accuracies of 91.86% and 94.37% under the 16-way 5-shot setting. The results highlight its potential as a practical solution for efficient new-subject enrollment in scenarios with limited labeled EEG data.
Tang et al. (Thu,) studied this question.