FederatedFactory: Generative One-Shot Learning for Extremely Non-IID Distributed Scenarios
This addresses the challenge of federated learning in medical imagery and other domains with highly skewed data distributions, offering a novel solution for data sovereignty without external dependencies.
The paper tackled the problem of federated learning failing in scenarios with extremely non-IID data by introducing FederatedFactory, a framework that uses generative priors instead of discriminative parameters, which recovered centralized performance and improved accuracy from 11.36% to 90.57% on CIFAR-10 under pathological heterogeneity.
Federated Learning (FL) enables distributed optimization without compromising data sovereignty. Yet, where local label distributions are mutually exclusive, standard weight aggregation fails due to conflicting optimization trajectories. Often, FL methods rely on pretrained foundation models, introducing unrealistic assumptions. We introduce FederatedFactory, a zero-dependency framework that inverts the unit of federation from discriminative parameters to generative priors. By exchanging generative modules in a single communication round, our architecture supports ex nihilo synthesis of universally class balanced datasets, eliminating gradient conflict and external prior bias entirely. Evaluations across diverse medical imagery benchmarks, including MedMNIST and ISIC2019, demonstrate that our approach recovers centralized upper-bound performance. Under pathological heterogeneity, it lifts baseline accuracy from a collapsed 11.36% to 90.57% on CIFAR-10 and restores ISIC2019 AUROC to 90.57%. Additionally, this framework facilitates exact modular unlearning through the deterministic deletion of specific generative modules.