mixEEG: Enhancing EEG Federated Learning for Cross-subject EEG Classification with Tailored mixup

TL;DR

This work tackles cross-subject EEG classification under privacy-preserving federated learning by introducing mixEEG, a framework that tailors mixup to EEG data and leverages averaged unlabeled target data under domain adaptation FL. It formulates DG FL and DA FL settings for FL with multiple clients and proposes three EEG-specific mixup strategies (Linear, Channel, Frequency) to boost transferability, demonstrating substantial gains over FedAvg on epilepsy detection and emotion recognition tasks with CNN and MLP backbones. Key contributions include the first investigation of DG/DA FL for cross-subject EEG, the mixEEG framework, and thorough ablations showing optimal mixup configurations and data-sharing parameters, all supporting privacy-preserving, scalable cross-subject BCI deployment. The findings highlight the practical impact of modality-aware augmentation and averaged data sharing on improving generalization in decentralized EEG systems.

Abstract

The cross-subject electroencephalography (EEG) classification exhibits great challenges due to the diversity of cognitive processes and physiological structures between different subjects. Modern EEG models are based on neural networks, demanding a large amount of data to achieve high performance and generalizability. However, privacy concerns associated with EEG pose significant limitations to data sharing between different hospitals and institutions, resulting in the lack of large dataset for most EEG tasks. Federated learning (FL) enables multiple decentralized clients to collaboratively train a global model without direct communication of raw data, thus preserving privacy. For the first time, we investigate the cross-subject EEG classification in the FL setting. In this paper, we propose a simple yet effective framework termed mixEEG. Specifically, we tailor the vanilla mixup considering the unique properties of the EEG modality. mixEEG shares the unlabeled averaged data of the unseen subject rather than simply sharing raw data under the domain adaptation setting, thus better preserving privacy and offering an averaged label as pseudo-label. Extensive experiments are conducted on an epilepsy detection and an emotion recognition dataset. The experimental result demonstrates that our mixEEG enhances the transferability of global model for cross-subject EEG classification consistently across different datasets and model architectures. Code is published at: https://github.com/XuanhaoLiu/mixEEG.

Figures (6)

• Figure 1: The problem setting of domain generalization federated learning (DG FL) and domain adaptation federated learning (DA FL), which aims to learn a global model from multiple decentralized source domains. The unlabeled sharing data is sent to client only under the DA FL setting.
• Figure 2: Three EEG mixup methods: (a) Linear Mixup, (b) Channel Mixup: an example that combines the left side and the right side of two EEG signals, (c) Frequency Mixup: an example that that combines the high frequency bands and the low frequency bands of two EEG signals.
• Figure 3: The accuracy, F1 score and Cohen’s Kappa score of LinearMixup with various $\alpha$ on the Epilepsy Detection dataset. The network architecture is MLP.
• Figure 4: The accuracy of the global model trained with different training set sizes on the SEED dataset. We choose the No.15 subject as the target domain, and adopt other 14 subjects as the training set. The network architecture is CNN.
• Figure 5: The 3D heatmap of the LOSO results on the SEED dataset with different sharing ratio $r$ and aggregating number $s$. The network architecture is MLP.