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

Conventional cross-subject electroencephalogram (EEG) signal identification frameworks require centralized aggregation of all subjects' data for feature extraction, which inherently poses substantial risks of data privacy breaches. In response to this critical issue, the present study delves into the classification of steady-state visual evoked potential (SSVEP) signals with an emphasis on data privacy preservation. First, we design a federated learning framework (FedGF) consisting of a central server and multiple clients, where the server generates global features and coordinates distributed training across clients, while retaining subject-specific raw data locally to ensure privacy protection. Then, to enhance model generalizability, FedGF employs data-free knowledge distillation (DFKD) to achieve knowledge transfer across clients through global feature learning. Extensive experiments on two public datasets (Dataset 1 'session01' and 2 'session02') and one private dataset (Dataset 3) demonstrate the superiority of the proposed method over baseline approaches, achieving performance improvements of 0.52%, 0.65%, and 0.53%, respectively. Finally, we develop a novel smart soft gripper with thermochromic capabilities and seamlessly integrate it with the trained network, demonstrating robust performance in daily grasping tasks. The source code is available at https://github.com/raow923/FedGF.