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

TL;DR

EEG-based motor imagery BCIs face privacy concerns when aggregating data across users. The paper introduces FedBS, a federated learning framework with local Batch-specific BN and a Sharpness-aware Minimization (SAM) optimizer to protect privacy while enhancing cross-subject generalization. Across three public MI datasets and three backbone models, FedBS outperforms six state-of-the-art FL methods and even centralized training, demonstrating a favorable privacy-utility trade-off. The approach enables scalable, privacy-preserving BCI training with improved decoding accuracy, suggesting strong practical impact for real-world BCIs.

Abstract

Training an accurate classifier for EEG-based brain-computer interface (BCI) requires EEG data from a large number of users, whereas protecting their data privacy is a critical consideration. Federated learning (FL) is a promising solution to this challenge. This paper proposes Federated classification with local Batch-specific batch normalization and Sharpness-aware minimization (FedBS) for privacy protection in EEG-based motor imagery (MI) classification. FedBS utilizes local batch-specific batch normalization to reduce data discrepancies among different clients, and sharpness-aware minimization optimizer in local training to improve model generalization. Experiments on three public MI datasets using three popular deep learning models demonstrated that FedBS outperformed six state-of-the-art FL approaches. Remarkably, it also outperformed centralized training, which does not consider privacy protection at all. In summary, FedBS protects user EEG data privacy, enabling multiple BCI users to participate in large-scale machine learning model training, which in turn improves the BCI decoding accuracy.

Figures (6)

• Figure 1: FedBS for privacy-preserving BCIs.
• Figure 2: Overview of FedBS. Each client represents an individual subject. 'Batch-specific BN' means the BN layer statistics are computed independently for each batch.
• Figure 3: Average classification accuracies of different FL approaches w.r.t. $P\cdot K$, the number of selected clients. GA was not included because it requires the participation of all clients in each round of training. (a) MI1; (b) MI2; and, (c) MI3.
• Figure 4: Average classification accuracies of different FL approaches w.r.t. $E$, the number of local computation epochs. (a) MI1; (b) MI2; and, (c) MI3.
• Figure 5: $t$-SNE visualization of feature extracted from test Subject 1 on the MI2 dataset. (a) CT; (b) FedAvg; and, (c) FedBS.