FederatedFactory: Generative One-Shot Learning for Extremely Non-IID Distributed Scenarios

Andrea Moleri; Christian Internò; Ali Raza; Markus Olhofer; David Klindt; Fabio Stella; Barbara Hammer

cs.LGMar 17, 2026

FederatedFactory: Generative One-Shot Learning for Extremely Non-IID Distributed Scenarios

Authors: Andrea Moleri, Christian Internò, Ali Raza, Markus Olhofer, David Klindt, Fabio Stella, Barbara Hammer

Organizations: Honda Research Institute Europe, Germany · University of Milan-Bicocca, Italy · Bielefeld University, Germany · Cold Spring Harbor Laboratory, U.S.

Abstract

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.

Figures & tables

Panel of Figure 1 — Figure 1: The Spectrum of Heterogeneity. (a) Ideal IID data (uniform overlap). (b) Dirichlet-distributed skew (imbalanced overlap). (c) Single-Class Silo (pathologically disjoint supports), representing the extreme theoretical limit.

Panel of Figure 2 — Figure 2: Centralized FederatedFactory Protocol. Aggregated Factories produce a fully synthetic dataset $\hat{\mathcal{D}}$ , enabling a global classifier training without raw data access.

Panel of Figure 3 — Figure 3: Decentralized FederatedFactory Protocol. This architecture depicts the P2P topology where local data flows ( $x\sim p_{k}(x)$ ) are augmented with synthetic samples from broadcasted Factories to train local experts, aggregated via PoE.

Panel of Figure 4 — Figure 4: Modular Unlearning Modes in the Generative Matrix $\boldsymbol{\Gamma}$ . By structuring the global model as a disjoint union of class-client generators $G_{\boldsymbol{\theta}_{c,k}}$ , FederatedFactory enables exact erasure without retraining the entire ensemble.

Method	CIFAR				BloodMNIST				PathMNIST
	Dirichlet		Silos		Dirichlet		Silos		Dirichlet		Silos
	Acc	AUC	Acc	AUC	Acc	AUC	Acc	AUC	Acc	AUC	Acc	AUC
FedAvg [ 36 ]	89.76 $\pm$ 1.12	99.22 $\pm$ 0.19	11.36 $\pm$ 1.28	50.91 $\pm$ 1.79	83.46 $\pm$ 5.18	97.57 $\pm$ 1.11	21.88 $\pm$ 5.39	55.23 $\pm$ 7.17	73.79 $\pm$ 16.86	90.11 $\pm$ 16.05	18.05 $\pm$ 0.80	50.12 $\pm$ 1.76
FedDyn [ 1 ]	61.82 $\pm$ 11.47	94.87 $\pm$ 2.65	10.12 $\pm$ 0.24	51.48 $\pm$ 0.83	64.93 $\pm$ 8.63	92.02 $\pm$ 2.21	19.47 $\pm$ 0.00	51.44 $\pm$ 2.87	60.99 $\pm$ 8.66	92.71 $\pm$ 4.49	17.76 $\pm$ 0.80	47.57 $\pm$ 3.22
FedProx [ 33 ]	89.67 $\pm$ 1.15	99.21 $\pm$ 0.19	18.08 $\pm$ 1.18	67.06 $\pm$ 1.02	84.11 $\pm$ 5.38	97.61 $\pm$ 1.00	20.18 $\pm$ 1.57	53.69 $\pm$ 3.32	77.09 $\pm$ 11.89	96.01 $\pm$ 4.44	18.93 $\pm$ 0.57	52.97 $\pm$ 2.67
Scaffold [ 27 ]	79.26 $\pm$ 5.10	97.73 $\pm$ 0.86	12.99 $\pm$ 0.75	54.83 $\pm$ 1.08	82.31 $\pm$ 2.97	97.51 $\pm$ 0.75	22.60 $\pm$ 1.96	67.84 $\pm$ 2.98	71.88 $\pm$ 9.70	95.61 $\pm$ 1.88	31.15 $\pm$ 2.10	64.87 $\pm$ 2.79

Table 1: Robustness to Extreme Statistical Heterogeneity. Mean

\pm

SD of Accuracy and AUROC (%) under moderate (

\alpha=0.1

) and pathological (

\alpha\rightarrow 0

) label skew. Best overall performance ( Red ). Bold indicates the best FL method.

Figure 5: Results in Pathological Heterogeneity. While standard baselines ( FedAvg

\circ

, FedDyn

\square

, FedProx

\triangle

, Scaffold

\diamond

) collapse as we move from moderate skew (

\alpha=0.1

) to extreme silos (

\alpha\rightarrow 0

), FederatedFactory matches the Upper Bound (—) in both Centralized (

+

) , Decentralized (

\star

) configurations.

Method	CIFAR		BloodMNIST		PathMNIST		RetinaMNIST		ISIC2019
Method	FLOPs	MBs	FLOPs	MBs	FLOPs	MBs	FLOPs	MBs	FLOPs	MBs
FedAvg [ 36 ]	$\mathbf{1.95\times 10^{17}}$	358,899.6	$\mathbf{3.57\times 10^{16}}$	287,069.6	$\mathbf{2.69\times 10^{17}}$	322,981.5	$\mathbf{3.22\times 10^{15}}$	179,371.6	$\mathbf{3.11\times 10^{17}}$	287,163.4
FedDyn [ 1 ]	$\mathbf{1.95\times 10^{17}}$	358,899.6	$\mathbf{3.57\times 10^{16}}$	287,069.6	$\mathbf{2.69\times 10^{17}}$	322,981.5	$\mathbf{3.22\times 10^{15}}$	179,371.6	$\mathbf{3.11\times 10^{17}}$	287,163.4
FedProx [ 33 ]	$\mathbf{1.95\times 10^{17}}$	358,899.6	$\mathbf{3.57\times 10^{16}}$	287,069.6	$\mathbf{2.69\times 10^{17}}$	322,981.5	$\mathbf{3.22\times 10^{15}}$	179,371.6	$\mathbf{3.11\times 10^{17}}$	287,163.4
Scaffold [ 27 ]	$\mathbf{1.95\times 10^{17}}$	717,799.1	$\mathbf{3.57\times 10^{16}}$	574,139.3	$\mathbf{2.69\times 10^{17}}$	645,962.9	$\mathbf{3.22\times 10^{15}}$	358,743.2	$\mathbf{3.11\times 10^{17}}$	574,326.8
FedFact (Server)	$7.30\times 10^{18}$	1,934.1	$4.47\times 10^{18}$	1,547.3	$5.03\times 10^{18}$	1,740.7	$2.79\times 10^{18}$	967.0	$2.52\times 10^{20}$	1,435.3

Table 2: Resource Trade-off Analysis. Total computational overhead (FLOPs) vs. communication volume (MBs) across the five benchmarks.

Figure 6: Real vs. Synthetic Data in Representation Space. Top: DINOv2 (Block 8) feature space [ 37 ] . FederatedFactory ( $\blacktriangle$ ) represents the true target manifolds data (•), preserving good performance. Bottom: Visual comparisons across datasets. Columns pair real and generated samples of the same class, demonstrating accurate morphological preservation without memorization.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Global Manifold Alignment in 2D t-SNE Subspace. Feature space projections comparing true localized training ( $\mathcal{D}_{\mathrm{union}}$ , $\circ$ ) and synthesized global ( $\hat{\mathcal{D}}_{\mathrm{syn}}$ , $\triangle$ ) distributions across all five benchmarks.

Figure 8: Quantitative Manifold Alignment via ECDF and Density Histograms. Evaluation of inter-class Fidelity (top rows: L2 distance from synthetic samples to their nearest real neighbors) and intra-class Diversity (bottom rows: L2 distance strictly among synthetic samples) across five benchmarks.

Figure 9: Absolute Performance Recovery under Single-Class Silos. Grouped bar chart detailing the accuracy and AUROC across the evaluated datasets. The gray shaded region represents the collapsed performance ceiling of standard iterative federated learning methods. FederatedFactory (in both Centralized and Decentralized modes) successfully escapes this collapsed regime, yielding improvements (e.g., matching the Centralized Upper Bound ).

Figure 10: The Compute-for-Bandwidth Swap. Log-log scatter comparing computational overhead versus network exposure. Iterative baselines ( FedAvg

\circ

, FedDyn

\square

, FedProx

\triangle

, Scaffold

\diamond

) cluster bottom-right (low compute, massive bandwidth). Conversely, FederatedFactory ( Centralized

+

, Decentralized

\star

) shifts to the top-left, accepting higher local FLOPs to achieve higher trustworthiness.

Dataset	Metric	Class
Dataset	Metric	0	1	2	3	4	5	6	7	8	9
CIFAR	FID ( $\downarrow$ )	29.07	22.50	27.28	46.67	22.10	35.40	41.02	15.01	19.32	25.75
CIFAR	KID ( $\downarrow$ )	1.71	1.38	1.59	3.48	1.50	2.16	3.09	0.58	1.00	1.85
BloodMNIST	FID ( $\downarrow$ )	84.01	61.54	45.08	50.37	60.58	66.42	50.78	27.19	–	–
BloodMNIST	KID ( $\downarrow$ )	10.76	8.83	4.63	6.33	7.54	7.69	6.76	2.97	–	–
RetinaMNIST	FID ( $\downarrow$ )	38.49	58.40	38.77	44.09	58.88	–	–	–	–	–

Appendix

Table 3: Per-Class Generative Alignment Metrics. Quantitative evaluation of the synthesized images via FID and KID (scaled by

10^{2}

). Lower values indicate better manifold alignment with the real target distribution. Values are rounded to two decimal places for readability. For CIFAR, classes 0–9 correspond to airplane, automobile, bird, cat, deer, dog, frog, horse, ship, and truck, respectively.

Figure 11: Quantitative Manifold Alignment via FID and KID. Per-class distributions of FID (left) and KID (right). The raincloud plots visualize the density, raw data points, and median scores.

Figure 12: Nearest Neighbor Analysis on CIFAR-10. Synthetic samples (top rows) paired with their closest real counterparts (bottom rows). Class 9 has been omitted to accommodate layout constraints.

Figure 13: Nearest Neighbor Analysis on BloodMNIST. The localized Factories successfully capture the distinct morphological features, shapes, and textures of different blood cell types.

Figure 14: Nearest Neighbor Analysis on PathMNIST. Synthetic histological patches alongside their closest real training samples.

Figure 15: Nearest Neighbor Analysis on RetinaMNIST. Despite the intrinsically low structural variance of retinal fundus crops, the generative models maintain sufficient diversity and do not trivially memorize the limited training support.

Figure 16: Nearest Neighbor Analysis on ISIC2019. High-resolution dermatoscopic lesions are synthesized with high fidelity.

Federated Learning

Federated learning trains shared models across data holders without requiring them to pool their raw data.

See all 951 papers

WHERE to Generate Matters: Budget-Aware Synthetic Augmentation for Label Skewed Federated Learning
Jul 7, 2026Sangwoo Lee, Sunghwan Park, Jaewoo LeeSynthetic Data Augmentation Data Augmentation
Chung-Ang University, Seoul, South Korea
GFPL: Generative Federated Prototype Learning for Resource-Constrained and Data-Imbalanced Vision Task
Feb 25, 2026Shiwei Lu, Yuhang He, Jiashuo Li +2Federated Learning Federated Learning Aggregation
Xi’an Jiaotong University · Air Force Engineering University
Asynchronous Federated Unlearning with Invariance Calibration for Medical Imaging
Apr 29, 2026Zhaoyuan Cai, Xinglin ZhangFederated Unlearning Medical Imaging
School of Computer Science and Engineering South China University of Technology Guangzhou, China

Jul 7, 2026J/K move · Enter open · S save

WHERE to Generate Matters: Budget-Aware Synthetic Augmentation for Label Skewed Federated Learning

Sangwoo Lee, Sunghwan Park, Jaewoo Lee

Chung-Ang University, Seoul, South Korea

Synthetic Data Augmentation Data Augmentation Non-IID Federated Learning Federated Learning

Open paper AskPDF arXiv

Label skew in federated learning (FL) causes client drift and degrades global accuracy. Synthetic data augmentation can reduce this imbalance; however, full class balancing requires substantial computation cost. We propose FedEAS, a policy that assigns each client an entropy-adaptive per-class generation budget computed from its local label distribution. The budget jointly decides how much each client generates and WHERE the samples go. Accordingly, the total generation budget follows from the per-client budgets rather than being fixed in advance. FedEAS recovers most of the accuracy gain of full class balancing while reducing the generation budget by 94.1%. At the same total generation budget, it outperforms Uniform allocation by up to 18.82% across CIFAR-10 and CIFAR-100.

Abstract

Federated Learning

Explore similar work

WHERE to Generate Matters: Budget-Aware Synthetic Augmentation for Label Skewed Federated Learning

GFPL: Generative Federated Prototype Learning for Resource-Constrained and Data-Imbalanced Vision Task

Asynchronous Federated Unlearning with Invariance Calibration for Medical Imaging