Alignment-Based Adversarial Training (ABAT) for Improving the Robustness and Accuracy of EEG-Based BCIs

Area of Science:

• Neuroscience
• Machine Learning
• Cybersecurity

Background:

• Machine learning significantly advances electroencephalogram (EEG) based brain-computer interfaces (BCIs).
• Existing BCI research primarily focuses on decoding accuracy, neglecting adversarial security vulnerabilities.
• Direct application of adversarial defenses from other domains degrades BCI classification accuracy on benign samples, limiting their utility.

Purpose of the Study:

• To address the challenge of applying adversarial defenses to EEG-based BCIs without compromising accuracy.
• To propose a novel method, alignment-based adversarial training (ABAT), for enhancing both the accuracy and robustness of EEG classifiers.
• To investigate the potential of adversarial attacks as a tool for improving BCI model performance.

Main Methods:

• Proposed ABAT, integrating EEG data alignment with adversarial training.
• Data alignment reduces distribution discrepancies between different EEG trial domains.
• Adversarial training strengthens the classification boundary for improved robustness.

Main Results:

• Demonstrated the effectiveness of ABAT across five EEG datasets and two BCI paradigms (motor imagery, event-related potential recognition).
• Validated ABAT with three convolutional neural network classifiers (EEGNet, ShallowCNN, DeepCNN) under diverse experimental settings.
• Showcased that adversarial attacks, when integrated into ABAT, simultaneously enhance model accuracy and robustness.

Conclusions:

• ABAT effectively improves EEG-based BCI accuracy and robustness.
• The proposed method offers a viable solution for securing BCIs against adversarial attacks.
• Adversarial training, paradoxically, can be leveraged to enhance BCI system performance.