DyAMNet: dynamic adversarial and contrastive network for EEG biometrics.

Introduction: Electroencephalogram (EEG)-based biometric recognition for brain-computer interfaces faces challenges from domain shifts, temporal nonstationarity, and limited scalability. Methods: To address these issues, we present DyAMNet, a framework that combines EEG microstate analysis with a hybrid attention mechanism. DyAMNet employs dynamic loss balancing to improve generalization and constructs a domain-invariant feature space that supports user expansion without catastrophic forgetting. We evaluated the model on three benchmark datasets (DEAP, THU-EP, and SEED). Results: The framework attains 87.2% accuracy in cross-dataset recognition and retains 84.0% accuracy when incrementally scaling to 60 users. The system also tolerates physiological artifacts and intersession signal drift, outperforming state-of-the-art models. Discussion: These findings show that dynamic adversarial training coupled with contrastive feature learning reduces brain-signal variability and preserves scalability. The work establishes a robust basis for feasible identity authentication and supports deploying brain-computer interfaces in clinical and everyday settings. The code is available at: https://github.com/cangtianhaoxue/DyAMNet.git.