Privacy-Preserving Domain Adaptation for Motor Imagery-Based Brain-Computer Interfaces.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Electroencephalogram (EEG) signals are crucial for motor imagery (MI) based brain-computer interfaces (BCIs).
• Domain adaptation techniques improve BCI accuracy for new users by leveraging data from existing users, but raise privacy concerns due to sensitive EEG data.
• Privacy-preserving domain adaptation is essential for secure and effective BCI applications.

Purpose of the Study:

• To develop a privacy-preserving domain adaptation method for EEG-based BCIs.
• To enhance classification accuracy for new users while protecting sensitive data from previous users.
• To address the challenge of source-free domain adaptation in the context of MI EEG.

Main Methods:

• Proposed Augmentation-based Source-Free Adaptation (ASFA) involving source and target model training.
• Introduced a novel data augmentation technique for MI EEG signals to boost cross-subject generalization.
• Incorporated uncertainty reduction and consistency regularization for robust target model training, requiring only source model parameters, not raw data.

Main Results:

• ASFA demonstrated superior performance compared to 15 classical and state-of-the-art MI classification methods.
• Experimental results on four MI datasets validated the effectiveness of the proposed ASFA approach.
• The method achieved high classification accuracy and robust privacy protection.

Conclusions:

• This work presents the first completely source-free domain adaptation for EEG-based BCIs.
• ASFA successfully balances high classification accuracy with strong privacy preservation.
• The findings are significant for the commercialization and widespread adoption of EEG-based BCIs.