Electroencephalography (EEG)-based motor imagery decoding plays an important role in brain-computer interface (BCI) systems. However, collecting and centralizing EEG data from multiple users raises significant privacy and data-sharing concerns. Federated learning provides a promising paradigm for collaborative model training across distributed clients without transferring raw neural recordings. Nevertheless, the high inter-subject variability of EEG signals often leads to unstable optimization and degraded performance in federated environments.  In this work, we propose a lightweight federated EEG classification framework that integrates Filter Bank Common Spatial Pattern (FBCSP) feature extraction with a distributed classifier trained using the Federated Averaging (FedAvg) algorithm. In the proposed pipeline, each subject is treated as an independent federated client, enabling collaborative learning while preserving data locality. Experiments were conducted on the EEG Motor Movement/Imagery dataset from PhysioNet involving 100 subjects performing left- and right-hand motor imagery tasks.  The results demonstrate that the proposed federated FBCSP framework achieves a classification accuracy of approximately 89.3%, significantly outperforming the baseline federated model which converges at around 81-82% accuracy. On average, the proposed method achieves a mean accuracy of 88.83%, compared with 80.95% for the baseline approach. In addition, the federated training process exhibits stable convergence, improving from approximately 79.9% accuracy in early communication rounds to 89.3% after 50 rounds. Our findings indicate that incorporating domain-specific spatial filtering significantly improves the robustness of federated EEG learning by reducing cross-subject variability before distributed optimization.  As a result, the proposed framework demonstrates that combining classical EEG signal processing techniques with federated learning provides an effective and privacy-preserving solution for large-scale collaborative BCI model training without requiring centralized access to sensitive neural data.

Brain–Computer Interface,Electroencephalography Signal,Filter Bank Common Spatial Pattern,Motor Imagery Classification,Privacy-Preserving Machine Learning