Federated Motor Imagery Classification for Privacy-Preserving Brain-Computer Interfaces.

Area of Science:

• Neuroscience
• Machine Learning
• Data Privacy

Background:

• Accurate electroencephalography (EEG)-based brain-computer interface (BCI) classification demands extensive user data.
• Protecting user data privacy is paramount in BCI research and development.
• Federated learning (FL) offers a decentralized approach to train models without centralizing sensitive data.

Purpose of the Study:

• To propose a novel federated learning framework, FedBS, for privacy-preserving EEG-based motor imagery (MI) classification.
• To enhance model generalization and reduce data discrepancies across diverse users in FL settings.
• To improve BCI decoding accuracy while safeguarding individual EEG data.

Main Methods:

• Federated classification with local Batch-specific batch normalization (BS) to mitigate data heterogeneity.
• Integration of Sharpness-aware minimization (SAM) optimizer for improved local model generalization.
• Experimental validation on three public MI datasets using diverse deep learning architectures.

Main Results:

• FedBS significantly outperformed six existing state-of-the-art FL methods in MI classification.
• The proposed FedBS approach surpassed the performance of centralized training, despite the latter not incorporating privacy measures.
• Demonstrated superior performance across multiple deep learning models and public datasets.

Conclusions:

• FedBS effectively protects user EEG data privacy during federated training.
• The framework enables collaborative training of BCI models with large-scale, multi-user EEG datasets.
• FedBS leads to enhanced BCI decoding accuracy through privacy-preserving machine learning.