Multi-Source Decentralized Transfer for Privacy-Preserving BCIs.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalogram (EEG) based brain-computer interfaces (BCIs) face challenges with high intra- and inter-subject variations.
• Existing transfer learning methods for BCIs often raise privacy concerns due to direct use of source subject data.

Purpose of the Study:

• To develop a decentralized, privacy-preserving transfer learning approach for EEG-based BCIs.
• To enable effective cross-subject transfer learning with unlabeled target EEG trials while protecting source subject data.

Main Methods:

• Proposed an offline unsupervised multi-source decentralized transfer (MSDT) approach.
• Generated pre-trained models from local source subject data.
• Performed decentralized transfer using either source model parameters (gray-box) or predictions (black-box).

Main Results:

• MSDT demonstrated superior performance compared to existing methods that do not prioritize privacy.
• Achieved high privacy-protection for source subjects' data.
• Showcased effectiveness across motor imagery and affective BCI paradigms.

Conclusions:

• The proposed MSDT approach effectively balances privacy preservation and classification performance in EEG BCIs.
• MSDT offers a viable solution for cross-subject transfer learning in decentralized BCI systems.
• This work advances the field of privacy-preserving machine learning in BCI applications.