An incremental adversarial training method enables timeliness and rapid new knowledge acquisition.

Area of Science:

• Neuroscience
• Computer Science
• Machine Learning

Background:

• Adversarial training defends deep models but is computationally expensive, hindering timely updates and new knowledge integration.
• Existing methods require full neural network retraining, posing challenges for dynamic applications like brain-computer interfaces (BCI).

Purpose of the Study:

• To propose and evaluate an incremental adversarial training (IncAT) method for deep learning models in BCI applications.
• To enhance model robustness against adversarial attacks while preserving performance on clean data and improving learning efficiency.

Main Methods:

• Developed a Neural Hybrid Assembly Network (NHANet) for BCI.
• Calculated the Fisher information matrix to identify important network parameters.
• Incorporated Elastic Weight Consolidation (EWC) loss during adversarial sample training to preserve critical parameters.

Main Results:

• Applied IncAT to the University of Bonn epilepsy BCI dataset.
• Achieved robust accuracies of 95.33% (FGSM), 94.67% (PGD), and 93.60% (BIM).
• Demonstrated significant accuracy improvements (5.06%, 4.67%, 2.67%) over traditional methods without impacting clean sample accuracy.

Conclusions:

• The proposed IncAT method offers an effective and computationally efficient defense against adversarial attacks in BCI.
• IncAT successfully enhances model robustness and generalization without sacrificing performance on clean data, validating its effectiveness and efficiency.