DyAMNet: dynamic adversarial and contrastive network for EEG biometrics.

Area of Science:

• Neuroscience
• Biometrics
• Machine Learning

Background:

• Electroencephalogram (EEG)-based biometric recognition for brain-computer interfaces (BCIs) faces significant challenges.
• These challenges include domain shifts, temporal nonstationarity, and limited scalability, hindering practical BCI deployment.

Purpose of the Study:

• To introduce DyAMNet, a novel framework designed to overcome the limitations of EEG-based biometric recognition.
• The framework aims to improve generalization, enable user expansion without catastrophic forgetting, and create a domain-invariant feature space.

Main Methods:

• DyAMNet integrates EEG microstate analysis with a hybrid attention mechanism.
• It employs dynamic loss balancing for improved generalization and contrastive feature learning to reduce brain-signal variability.
• The framework was evaluated on three benchmark datasets: DEAP, THU-EP, and SEED.

Main Results:

• DyAMNet achieved 87.2% accuracy in cross-dataset recognition.
• It maintained 84.0% accuracy when incrementally scaling to 60 users.
• The system demonstrated tolerance to physiological artifacts and intersession signal drift, outperforming existing state-of-the-art models.

Conclusions:

• Dynamic adversarial training and contrastive feature learning effectively reduce brain-signal variability and enhance scalability.
• The DyAMNet framework provides a robust foundation for reliable identity authentication.
• This work supports the deployment of BCIs in both clinical and everyday settings.