A Robustness Study of Federated Learning for EEG-Based Eye State Classification

Electroencephalogram (EEG) data is crucial for understanding brain activities, providing insights into cognitive functions. Due to their sensitive nature, EEG data analysis requires robust privacy protection measures. Federated learning (FL) addresses these concerns by keeping data on local devices (clients) and sharing only model updates with the central server, thus enhancing privacy compared to traditional centralized learning. This paper focuses on utilizing EEG data for eye state classification while preserving privacy. The baseline accuracy and recall of the EEG-based eye state classification is established through centralized learning while the privacy concern is addressed using federated learning (FL). Our initial results indicate that while FL could preserve more privacy, it leads to a noticeable reduction in both accuracy and recall compared to the centralized learning approach. To enhance the FL model performance, we explore various configurations of EEG-based eye state classification using FL. These include adjustments to learning rate, data diversity, and aggregation algorithms. We found that with a learning rate of 1e-5 and the FedProx aggregation algorithm, our FL model achieved accuracy and recall comparable to the centralized learning model, especially as the data diversity at each client site increased. Additionally, we evaluated the robustness of our FL model against imbalanced data distributions. The results show that our FL model effectively maintains its good performance even with imbalanced data distributions among classes. Lastly, we conducted a comprehensive evaluation of the computational and communication trade-offs. Our results demonstrate that federated learning is a practical solution for resource-constrained environments, enabling efficient decentralized EEG-based eye state classification systems.