FedVR: Variance Regularized Hypernetwork for Federated Domain Generalization

Sunny Gupta, Shambhavi Shanker, Amit Sethi

Keywords: Federated Learning, Domain Generalization, Hypernetwork, Differential Privacy

TL;DR: FedVR introduces a variance-regularized hypernetwork that generates personalized client models and achieves robust domain generalization under non-IID federated settings.

Abstract:

Federated Learning (FL) enables model training across decentralized and privacy-sensitive data sources, but its effectiveness is severely degraded by domain shifts. Conventional domain generalization methods aim to extract invariant features, yet their integration into FL is hindered by privacy constraints and the limitations of linear aggregation in FedAvg. Hypernetwork-based approaches offer non-linear parameter synthesis, but existing methods lack explicit fairness guarantees and remain fragile under strong heterogeneity. We propose FedVR, a framework for hypernetwork-driven generalization adjustment in federated domain generalization. FedVR generates per-domain models via a hypernetwork conditioned on distributional embeddings that summarize each client’s data statistics. To improve robustness, we extend generalization adjustment to these hypernetwork-generated models, explicitly minimizing the variance of per-domain generalization gaps. Moreover, FedVR enables zero-shot deployment on unseen domains: given a small unlabeled or lightly labeled pilot set, the domain encoder produces a new embedding from which the hypernetwork synthesizes specialized parameters without fine-tuning.
Extensive experiments on PACS, Office-Home, and VLCS benchmarks demonstrate that FedVR achieves superior accuracy, fairness, and calibration across both in-domain and out-of-domain settings, outperforming strong federated baselines.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 22