Adversarial artifact detection in EEG-based brain-computer interfaces.

Area of Science:

• Neuroscience
• Machine Learning
• Computer Science

Background:

• Machine learning has advanced electroencephalogram (EEG) based brain-computer interfaces (BCIs), but these systems are vulnerable to adversarial attacks.
• Adversarial examples, created by subtle input perturbations, can cause misclassifications, posing a security risk.
• Detecting these adversarial examples is vital for understanding BCI vulnerabilities and developing robust defenses.

Purpose of the Study:

• This research is the first to explore adversarial detection specifically within EEG-based BCIs.
• The study aims to evaluate and propose new methods for identifying adversarial examples in BCI systems.
• Key objectives include assessing detection performance against various attacks and understanding the transferability of detectors.

Main Methods:

• Adapted popular adversarial detection techniques from computer vision for BCI applications.
• Developed two novel Mahalanobis distance-based detection methods.
• Proposed three new cosine distance-based detection approaches.
• Evaluated eight detection methods across three EEG datasets, three neural networks, and four adversarial attack types.

Main Results:

• Achieved an Area Under the Curve (AUC) score of up to 99.99% in detecting white-box adversarial attacks.
• Demonstrated promising performance of the proposed Mahalanobis and cosine distance-based detectors.
• Assessed the transferability of adversarial detectors against unknown attack types, revealing distinct distributions among adversarial examples.

Conclusions:

• White-box adversarial attacks on EEG-BCIs can be effectively detected.
• Differences in adversarial example distributions necessitate tailored detection strategies.
• This work provides foundational insights for enhancing the security and reliability of future BCI models.