FederatedFactory: Generative One-Shot FL

• FederatedFactory is a federated learning framework that inverts traditional methods by transmitting local generative modules instead of discriminative updates.
• It overcomes non-IID challenges by synthesizing a globally balanced dataset from single-class silos without relying on external pretrained models.
• The framework offers modular unlearning and significant communication efficiencies, achieving near-centralized performance across diverse benchmarks.

FederatedFactory is a federated learning framework introduced for extremely non-IID distributed scenarios, especially the single-class silo regime in which clients have mutually exclusive label supports and standard weight aggregation collapses. Its defining move is to invert the unit of federation from discriminative model parameters to locally trained generative priors: each client transmits a generator in a single communication round, the server synthesizes a globally class-balanced dataset from the resulting module bank, and a downstream classifier is trained on synthetic data rather than by iterative averaging of classifier updates (Moleri et al., 17 Mar 2026). In adjacent manufacturing literature, the term also supports a broader reading as a cross-factory collaboration pattern in which organizations improve shared AI services without exchanging raw data; that broader interpretation is suggested by earlier federated quality-inspection work rather than formally standardized as a single platform architecture (Hegiste et al., 2023).

1. Problem regime and conceptual shift

FederatedFactory targets a pathological heterogeneity regime in cross-silo federated learning: clients are persistent institutions, but their local label supports can be disjoint. In the strongest case, each client contains only one class, so local discriminative optimization cannot learn meaningful inter-class boundaries. The framework starts from the standard federated objective

$$\min_{\mathbf{w}} \mathcal{L}(\mathbf{w}) := \frac{1}{K}\sum_{k=1}^{K}\mathcal{L}_k(\mathbf{w}), \qquad \mathcal{L}_k(\mathbf{w}) = \mathbb{E}_{(x,y)\sim p_k}[\ell(\mathbf{w};x,y)],$$

but focuses on the limit in which

$$\mathcal{Y}_i \cap \mathcal{Y}_j = \emptyset \quad \text{for } i\neq j.$$

Under this condition, the paper argues that local optimization trajectories become actively conflicting, so FedAvg-style aggregation can collapse to near-random or majority-class behavior (Moleri et al., 17 Mar 2026).

The framework is explicitly designed to satisfy four constraints in combination: pathological skew, zero dependency on pretrained foundation models or public proxy data, strict data sovereignty, and one-shot communication. The zero-dependency claim is central. Many one-shot or synthetic-data federated methods use external priors such as CLIP, Stable Diffusion, or other foundation models. FederatedFactory rejects that assumption, particularly for medical imaging, on the grounds that external priors can fail to represent rare or domain-specific structure and may erase diagnostically relevant off-manifold features (Moleri et al., 17 Mar 2026).

The conceptual shift is therefore not a new server optimizer but a change in what is federated. Standard federated learning shares discriminative updates. FederatedFactory shares generative modules. Each client trains a local generator on its private distribution, and the server later constructs a global training distribution by sampling from the collection of client generators. This removes iterative gradient conflict and avoids dependence on external priors. The paper calls the resulting server-side synthesis “ex nihilo synthesis,” meaning that the server creates a usable global training set without receiving raw client data and without relying on a pretrained external generator (Moleri et al., 17 Mar 2026).

2. Generative one-shot architecture

In the centralized protocol, each client $k$ first trains a local generator $G_{\boldsymbol{\theta}_k}$ on its private dataset $\mathcal{D}_k$. In the reported implementation, these generators are EDM2 diffusion models, although the framework is described as architecture-agnostic and could in principle use VAEs or GANs. After local training, each client sends the generator parameters $\boldsymbol{\theta}_k$ to the server exactly once. The server collects the generators into a library

$$\boldsymbol{\Theta} = \{G_{\boldsymbol{\theta}_1}, \dots, G_{\boldsymbol{\theta}_K}\},$$

samples latent noise

$$\mathbf{z}\sim\mathcal{N}(\mathbf{0},\mathbf{I}), \qquad \hat{\mathbf{x}} = G_{\boldsymbol{\theta}_k}(\mathbf{z}),$$

and labels each generated sample with the class represented by that module. The union of these samples forms a synthetic global dataset on which the server trains a discriminative classifier, implemented as ResNet-50 in the experiments (Moleri et al., 17 Mar 2026).

The paper also defines a more general class-by-client “Generative Matrix”

$\boldsymbol{\Gamma}\in\mathcal{F}^{C\times K},$

whose entry $\boldsymbol{\Gamma}_{c,k}$ contains a class-specific generator $$\mathcal{Y}_i \cap \mathcal{Y}_j = \emptyset \quad \text{for } i\neq j.$$0 or $$\mathcal{Y}_i \cap \mathcal{Y}_j = \emptyset \quad \text{for } i\neq j.$$1 if client $$\mathcal{Y}_i \cap \mathcal{Y}_j = \emptyset \quad \text{for } i\neq j.$$2 lacks class $$\mathcal{Y}_i \cap \mathcal{Y}_j = \emptyset \quad \text{for } i\neq j.$$3. When multiple clients contain the same class, server-side synthesis uses a quota rule based on local sample counts: $$\mathcal{Y}_i \cap \mathcal{Y}_j = \emptyset \quad \text{for } i\neq j.$$4 where $$\mathcal{Y}_i \cap \mathcal{Y}_j = \emptyset \quad \text{for } i\neq j.$$5 is the number of class-$$\mathcal{Y}_i \cap \mathcal{Y}_j = \emptyset \quad \text{for } i\neq j.$$6 samples at client $$\mathcal{Y}_i \cap \mathcal{Y}_j = \emptyset \quad \text{for } i\neq j.$$7, $$\mathcal{Y}_i \cap \mathcal{Y}_j = \emptyset \quad \text{for } i\neq j.$$8 is the set of clients containing class $$\mathcal{Y}_i \cap \mathcal{Y}_j = \emptyset \quad \text{for } i\neq j.$$9, and $k$0 is the target class size. This keeps synthesis in data space rather than by averaging generator parameters (Moleri et al., 17 Mar 2026).

A decentralized peer-to-peer variant is also defined. In that protocol, each client broadcasts its generator parameters to all peers, builds a mixed local dataset consisting of its own real data plus synthetic samples from other clients’ generators, trains a local expert classifier, and aggregates predictions at inference time through a Product of Experts: $k$1 with implementation in log-space and a minimum probability floor to avoid numerical veto by a single miscalibrated expert. This P2P variant is the stronger empirical configuration on several benchmarks (Moleri et al., 17 Mar 2026).

3. Mathematical formulation and theoretical claims

The theoretical analysis treats federated aggregation as a low-rank generative reconstruction problem. Under disjoint label support, the inaccessible true global joint distribution is written as

$k$2

with mixture weights

$k$3

After clients transmit generators, the server induces a synthetic global distribution

$k$4

The analysis assumes local generative convergence,

$k$5

and bounded loss,

$k$6

Under these assumptions, the paper proves a “Global Manifold Recovery” lemma: $k$7 Defining

$k$8

and letting $k$9 minimize synthetic risk while $G_{\boldsymbol{\theta}_k}$0 minimizes true risk, the main “Zero-Dependency Aggregation” theorem states

$G_{\boldsymbol{\theta}_k}$1

The bound is obtained by combining the KL lemma with Pinsker’s inequality and a total-variation bound on risk mismatch (Moleri et al., 17 Mar 2026).

The theoretical contribution is deliberately narrow. It does not supply a convergence theorem for iterative discriminative optimization, because the framework avoids iterative discriminative aggregation. It also does not claim universal recovery of centralized performance on every benchmark. Instead, it formalizes a conditional statement: if local generators approximate their private class-conditionals well, then a classifier trained on the synthetic federation distribution is close to the centralized optimum. The paper presents this as a contrast with foundation-model-dependent methods, whose error may be dominated by an uncontrollable projection or out-of-distribution term (Moleri et al., 17 Mar 2026).

4. Empirical behavior under pathological heterogeneity

The evaluation uses five image-classification datasets: CIFAR-10, BloodMNIST, PathMNIST, RetinaMNIST, and ISIC2019. Baselines are FedAvg, FedDyn, FedProx, SCAFFOLD, FederatedFactory centralized, FederatedFactory P2P, and a centralized upper bound. The classifier backbone is ResNet-50 trained with SGD, batch size $G_{\boldsymbol{\theta}_k}$2, weight decay $G_{\boldsymbol{\theta}_k}$3, initial learning rate $G_{\boldsymbol{\theta}_k}$4, cosine annealing, and mean $G_{\boldsymbol{\theta}_k}$5 standard deviation over five random seeds. The baselines use 200 communication rounds with 5 local epochs per round, while synthetic-data training uses 300 epochs. The key test regime is single-class silos, corresponding to $G_{\boldsymbol{\theta}_k}$6 in the Dirichlet-skew framing (Moleri et al., 17 Mar 2026).

On CIFAR-10 under single-class silos, the collapse of standard federated averaging is explicit. FedAvg reaches $G_{\boldsymbol{\theta}_k}$7 accuracy and $G_{\boldsymbol{\theta}_k}$8 AUC; FedDyn reaches $G_{\boldsymbol{\theta}_k}$9 accuracy and $\mathcal{D}_k$0 AUC; FedProx reaches $\mathcal{D}_k$1 accuracy and $\mathcal{D}_k$2 AUC; SCAFFOLD reaches $\mathcal{D}_k$3 accuracy and $\mathcal{D}_k$4 AUC. In the same regime, FederatedFactory reaches $\mathcal{D}_k$5 accuracy and $\mathcal{D}_k$6 AUC in the centralized variant, and $\mathcal{D}_k$7 accuracy and $\mathcal{D}_k$8 AUC in the P2P variant, against a centralized upper bound of $\mathcal{D}_k$9 accuracy and $\boldsymbol{\theta}_k$0 AUC (Moleri et al., 17 Mar 2026).

The medical benchmarks are more mixed but remain central to the paper’s claim. On BloodMNIST in the silo regime, FederatedFactory centralized reaches $\boldsymbol{\theta}_k$1 accuracy and $\boldsymbol{\theta}_k$2 AUC, compared with a centralized upper bound of $\boldsymbol{\theta}_k$3 accuracy and $\boldsymbol{\theta}_k$4 AUC. On ISIC2019, the P2P variant reaches $\boldsymbol{\theta}_k$5 accuracy and $\boldsymbol{\theta}_k$6 AUC, while the centralized upper bound is $\boldsymbol{\theta}_k$7 accuracy and $\boldsymbol{\theta}_k$8 AUC. PathMNIST is notably weaker: FederatedFactory centralized reaches $\boldsymbol{\theta}_k$9 accuracy and $$\boldsymbol{\Theta} = \{G_{\boldsymbol{\theta}_1}, \dots, G_{\boldsymbol{\theta}_K}\},$$0 AUC, against a centralized upper bound of $$\boldsymbol{\Theta} = \{G_{\boldsymbol{\theta}_1}, \dots, G_{\boldsymbol{\theta}_K}\},$$1 accuracy and $$\boldsymbol{\Theta} = \{G_{\boldsymbol{\theta}_1}, \dots, G_{\boldsymbol{\theta}_K}\},$$2 AUC. RetinaMNIST is unusual in the opposite direction: the P2P variant reaches $$\boldsymbol{\Theta} = \{G_{\boldsymbol{\theta}_1}, \dots, G_{\boldsymbol{\theta}_K}\},$$3 accuracy and $$\boldsymbol{\Theta} = \{G_{\boldsymbol{\theta}_1}, \dots, G_{\boldsymbol{\theta}_K}\},$$4 AUC, exceeding the centralized upper bound of $$\boldsymbol{\Theta} = \{G_{\boldsymbol{\theta}_1}, \dots, G_{\boldsymbol{\theta}_K}\},$$5 accuracy and $$\boldsymbol{\Theta} = \{G_{\boldsymbol{\theta}_1}, \dots, G_{\boldsymbol{\theta}_K}\},$$6 AUC (Moleri et al., 17 Mar 2026).

The empirical case is reinforced by generative-quality diagnostics rather than by classifier scores alone. The paper reports t-SNE and DINOv2 feature-space alignment between real and synthetic distributions, nearest-neighbor comparisons against memorization, ECDF and distance histograms for fidelity and diversity, and classwise FID/KID analysis. The reported CIFAR classwise FID is roughly around a mean of $$\boldsymbol{\Theta} = \{G_{\boldsymbol{\theta}_1}, \dots, G_{\boldsymbol{\theta}_K}\},$$7; KID values are generally $$\boldsymbol{\Theta} = \{G_{\boldsymbol{\theta}_1}, \dots, G_{\boldsymbol{\theta}_K}\},$$8 when scaled by $$\boldsymbol{\Theta} = \{G_{\boldsymbol{\theta}_1}, \dots, G_{\boldsymbol{\theta}_K}\},$$9 in the appendix table; and ISIC2019 median FIDs are around $$\mathbf{z}\sim\mathcal{N}(\mathbf{0},\mathbf{I}), \qquad \hat{\mathbf{x}} = G_{\boldsymbol{\theta}_k}(\mathbf{z}),$$0 with low KID values. The paper interprets these results as evidence that the generators cover the target manifolds well enough for downstream classification (Moleri et al., 17 Mar 2026).

5. Communication profile, modular unlearning, and limitations

FederatedFactory is explicitly positioned as one-shot communication. On CIFAR-10, the paper reports that FedAvg, FedDyn, and FedProx require $$\mathbf{z}\sim\mathcal{N}(\mathbf{0},\mathbf{I}), \qquad \hat{\mathbf{x}} = G_{\boldsymbol{\theta}_k}(\mathbf{z}),$$1 MB of communication, SCAFFOLD requires $$\mathbf{z}\sim\mathcal{N}(\mathbf{0},\mathbf{I}), \qquad \hat{\mathbf{x}} = G_{\boldsymbol{\theta}_k}(\mathbf{z}),$$2 MB, FederatedFactory centralized requires $$\mathbf{z}\sim\mathcal{N}(\mathbf{0},\mathbf{I}), \qquad \hat{\mathbf{x}} = G_{\boldsymbol{\theta}_k}(\mathbf{z}),$$3 MB, and the decentralized P2P variant requires $$\mathbf{z}\sim\mathcal{N}(\mathbf{0},\mathbf{I}), \qquad \hat{\mathbf{x}} = G_{\boldsymbol{\theta}_k}(\mathbf{z}),$$4 MB. The centralized version is therefore reported as a $$\mathbf{z}\sim\mathcal{N}(\mathbf{0},\mathbf{I}), \qquad \hat{\mathbf{x}} = G_{\boldsymbol{\theta}_k}(\mathbf{z}),$$5 communication reduction relative to the iterative baselines. The corresponding trade-off is a shift from a bandwidth-bound to a compute-bound regime: on CIFAR-10, FedAvg-like methods require $$\mathbf{z}\sim\mathcal{N}(\mathbf{0},\mathbf{I}), \qquad \hat{\mathbf{x}} = G_{\boldsymbol{\theta}_k}(\mathbf{z}),$$6 FLOPs, whereas FederatedFactory centralized requires $$\mathbf{z}\sim\mathcal{N}(\mathbf{0},\mathbf{I}), \qquad \hat{\mathbf{x}} = G_{\boldsymbol{\theta}_k}(\mathbf{z}),$$7 FLOPs and FederatedFactory P2P requires $$\mathbf{z}\sim\mathcal{N}(\mathbf{0},\mathbf{I}), \qquad \hat{\mathbf{x}} = G_{\boldsymbol{\theta}_k}(\mathbf{z}),$$8 FLOPs (Moleri et al., 17 Mar 2026).

A second distinctive feature is modular unlearning. Because the shared knowledge base is a set of separate generators rather than an entangled discriminative model, deletion is performed structurally. In the simple single-class case, forgetting client $$\mathbf{z}\sim\mathcal{N}(\mathbf{0},\mathbf{I}), \qquad \hat{\mathbf{x}} = G_{\boldsymbol{\theta}_k}(\mathbf{z}),$$9 means

$\boldsymbol{\Gamma}\in\mathcal{F}^{C\times K},$0

In the more general Generative Matrix formulation, the paper defines three deletion modes: vertical unlearning, $\boldsymbol{\Gamma}\in\mathcal{F}^{C\times K},$1, which forgets all modules from client $\boldsymbol{\Gamma}\in\mathcal{F}^{C\times K},$2; horizontal unlearning, $\boldsymbol{\Gamma}\in\mathcal{F}^{C\times K},$3, which removes class $\boldsymbol{\Gamma}\in\mathcal{F}^{C\times K},$4 globally; and targeted unlearning, $\boldsymbol{\Gamma}\in\mathcal{F}^{C\times K},$5, which removes a specific class from a specific client. After deletion, affected synthetic data are flushed and the classifier is retrained. The claim of “exact” unlearning is therefore structural and operational: the deleted generator is no longer a source of future synthesis (Moleri et al., 17 Mar 2026).

The limitations are equally explicit. Generator quality is critical, since the theory depends on bounded local generative error. The framework provides no formal privacy guarantee such as differential privacy, and the paper explicitly acknowledges the possibility of extraction and membership-inference risks in diffusion models. The evaluation scope is image classification only. Scalability may become problematic if every class-client pair requires its own generator in large heterogeneous systems. The cleanest theory is built around disjoint class supports, even though the paper sketches a broader multi-class extension. Finally, the empirical record is uneven: the PathMNIST gap shows that near-centralized recovery is not universal (Moleri et al., 17 Mar 2026).

6. Manufacturing interpretations and adjacent systems

A broader industrial interpretation of FederatedFactory is already visible in earlier manufacturing work. The paper “Federated Object Detection for Quality Inspection in Shared Production” presents a concrete early step toward what the term would mean in practice: multiple factories collaborate on a quality-inspection model without exchanging raw images, despite non-IID client distributions and heterogeneous backgrounds, lighting, and product variants. In its cabin-inspection experiment on unseen cross-client combinations, the global federated model reached $\boldsymbol{\Gamma}\in\mathcal{F}^{C\times K},$6 and $\boldsymbol{\Gamma}\in\mathcal{F}^{C\times K},$7, whereas the local client models achieved $\boldsymbol{\Gamma}\in\mathcal{F}^{C\times K},$8 and $\boldsymbol{\Gamma}\in\mathcal{F}^{C\times K},$9 on the same held-out setting. That work uses YOLOv5 with FedAvg rather than generative one-shot synthesis, but it already frames the central cross-factory problem: no single site has the full production distribution, yet a shared model should generalize across all variants while preserving raw-data privacy (Hegiste et al., 2023).

The same manufacturing paper also pushes beyond training into service integration. It connects federated inspection to “Production Level 4 Shared Production,” proposes offering the inspection capability as Software-as-a-Service via a marketplace, and discusses Asset Administration Shell self-description, Gaia-X connectors, service catalogs, downloadable Docker containers, and the possibility that customers update only model weights from a global federated model if the service is already deployed on-site. This broader service interpretation suggests that FederatedFactory, as a platform concept, is not exhausted by a single learning algorithm. It also involves onboarding rules, label governance, trustworthy collaboration, and interoperable deployment artifacts (Hegiste et al., 2023).

Adjacent systems literature suggests two further architectural layers for such a platform. A production-oriented runtime can be seen in Fed-DART and FACT, which separate a server-centric federated runtime and orchestration layer from a higher-level FL toolkit supporting aggregation and clustering in centralized, horizontal, cross-silo settings (Weber et al., 2022). A sovereignty-aware admission layer can be seen in FCaC, which introduces Virtual Federated Platforms and a cryptographic trust chain based on KYO, ECTs, and proof of possession, so that boundary admission becomes local verification rather than online policy evaluation (Fenoglio et al., 18 Mar 2026). This suggests, by implication rather than by explicit standardization, that a fuller FederatedFactory stack could combine manufacturing-specific model collaboration, production-grade orchestration, and cryptographically verifiable cross-boundary governance.