Fed-VQCS: Vector-Quantized Content-Style State-Space Disentanglement for Federated EEG Generalization

Federated learning provides a natural framework for decentralized EEG analysis, but learning shared and transferable representations remains challenging because task-relevant neural patterns are often entangled with subject-dependent characteristics in local EEG data. Existing federated EEG methods mainly focus on backbone design, aggregation strategies, or personalized optimization, while representation-level control of what information is globally shared remains insufficiently explored. In this work, we propose Fed-VQCS, a vector-quantized content-style state-space disentanglement framework for robust federated EEG learning. Fed-VQCS introduces a plug-and-play Vector-Quantized Content-Style State-Space (VQCS) module that can be inserted after different EEG feature extractors. Given a latent EEG representation, a bidirectional state-space semantic extractor first produces a continuous task-semantic representation, which is further mapped to globally shared discrete content prototypes through vector quantization. A complementary residual pathway captures variation outside the continuous semantic estimate and is adversarially constrained to suppress task-discriminative information. During federated training, the shared model parameters and content codebook are aggregated globally, while the residual pathway is evaluated within each client’s local training pass. Experiments on BCI Competition IV-2a and PhysioNet, repeated over five random seeds, show consistent improvements in accuracy and Cohen’s κ across four EEG backbones. Comparisons with representative federated learning methods, sensitivity analyses, ablation studies, low-data adaptation, and representation probing further support the effectiveness and robustness of the proposed framework under heterogeneous subject distributions.