Prediction Consistency and Confidence-Based Proxy Domain Construction for Privacy-Preserving in Cross-Subject EEG...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Domain adaptation is crucial for cross-subject electroencephalography (EEG) classification, addressing inter-subject variability.
• Existing methods often require direct access to source domain data, posing privacy concerns.
• Unlabeled target domain data is common in real-world EEG applications.

Purpose of the Study:

• To propose a privacy-preserving domain adaptation framework for cross-subject EEG classification.
• To develop a method that transfers knowledge from source to target domains without accessing raw source data.
• To mitigate the inter-subject variability problem in EEG analysis.

Main Methods:

• Proposed a novel framework, Prediction Consistency and Confidence (PDCC), to construct a proxy domain.
• The proxy domain, derived from source models' predictions on target data, replaces direct source data access.
• Employed decentralized training of source models and data augmentation/alignment for enhanced generalizability.

Main Results:

• PDCC effectively transfers knowledge from source to target domains while preserving source data privacy.
• Experimental results on four benchmark EEG datasets show PDCC outperforms eleven existing methods.
• The effectiveness of the proxy domain construction was extensively validated.

Conclusions:

• PDCC offers a robust and privacy-preserving solution for domain adaptation in cross-subject EEG classification.
• The proposed proxy domain effectively encapsulates source knowledge without compromising data privacy.
• This method significantly advances the applicability of EEG-based BCI systems in real-world scenarios.