Improving Generalization in Federated Learning for Steady-State Visual Evoked Potential Classification and Its...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Conventional electroencephalogram (EEG) analysis requires centralizing sensitive subject data, posing privacy risks.
• Classifying steady-state visual evoked potential (SSVEP) signals is crucial for brain-computer interfaces but faces privacy challenges.

Purpose of the Study:

• To develop a privacy-preserving framework for cross-subject SSVEP signal classification.
• To enhance model generalizability and performance using federated learning and knowledge distillation.

Main Methods:

• Designed a federated learning framework (FedGF) with a central server and local clients for distributed EEG training.
• Implemented data-free knowledge distillation (DFKD) for cross-client knowledge transfer via global feature learning.
• Validated the framework on public and private SSVEP datasets.

Main Results:

• FedGF demonstrated superior performance over baseline methods on three datasets, with accuracy improvements of 0.52%, 0.65%, and 0.53%.
• The framework successfully preserved subject-specific data privacy by retaining raw data locally.
• Achieved enhanced model generalizability through federated learning and DFKD.

Conclusions:

• The proposed FedGF framework offers an effective and privacy-preserving solution for cross-subject SSVEP classification.
• The integration with a novel smart soft gripper demonstrates the practical applicability of the trained network.
• The study highlights the potential of federated learning in sensitive biomedical data analysis.