Data alignment based adversarial defense benchmark for EEG-based BCIs.

Area of Science:

• Neuroscience
• Computer Science
• Artificial Intelligence

Background:

• Machine learning is widely used for signal decoding in electroencephalogram (EEG)-based brain-computer interfaces (BCIs).
• Existing research primarily focuses on improving BCI accuracy, often neglecting crucial security aspects.
• Recent studies highlight the vulnerability of EEG-based BCIs to sophisticated adversarial attacks.

Purpose of the Study:

• To establish the first adversarial defense benchmark for EEG-based BCIs, emphasizing data alignment.
• To enhance both the accuracy and robustness of EEG-based BCIs against potential threats.
• To provide comprehensive insights into the effectiveness of various defense strategies.

Main Methods:

• Evaluation of nine adversarial defense approaches, including five distinct defense strategies.
• Testing across five diverse EEG datasets, encompassing three experimental paradigms.
• Analysis using three different neural network architectures and four experimental scenarios.

Main Results:

• The integration of data augmentation, data alignment, and robust training demonstrably improves BCI accuracy and robustness.
• This combined approach surpasses the performance of using only one or two of these techniques.
• Detailed insights into the performance characteristics of EEG data alignment-based defenses were elucidated.

Conclusions:

• A combined strategy of data augmentation, data alignment, and robust training offers superior performance for EEG-based BCIs.
• The developed benchmark provides valuable guidance for creating more accurate and secure EEG-based BCIs.
• Addressing adversarial vulnerabilities is critical for the future development and deployment of BCI technology.