Research map / Federated learning and privacy / Personalized and heterogeneous FL
One-Shot Heterogeneous Federated Learning with Local Model-Guided Diffusion Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FedLMG, a heterogeneous one-shot federated learning method where clients train only local classifiers and the server uses them (via CE and BN losses) to guide a pretrained Stable Diffusion model in generating synthetic data for aggregation.
Problem
One-shot federated learning with diffusion models typically requires foundation models (CLIP, BLIP, or client-side diffusion training) on client devices, increasing client compute and limiting heterogeneous client model support.
Approach
Clients upload only their trained classifiers; the server guides a diffusion model's sampling with the classifier's cross-entropy loss on predicted x0 plus a BN-statistics matching loss, producing labeled synthetic data that matches client distributions for training the aggregated model via fine-tuning or distillation.
Key results
- FedLMG_SD reaches 58.86% average accuracy on OpenImage versus 53.38% for FGL and 56.03% for the centralized ceiling.
- FedLMG_SD reaches 58.16% average on Common NICO++ versus 54.79% for FGL and 60.06% for the ceiling.
- Client FLOPs are 3.64G for FedLMG versus 227.34G for FGL and 334.73G for FedDISC; FedLMG has no download cost.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedLMG with specific-teacher distillation instead of FGL | FedLMG on OpenImage · OpenImage feature distribution skew | accuracy (Avg): 53.38 -> 58.86 | Helped |
| Use FedLMG with multi-teacher distillation instead of FGL | FedLMG on DomainNet · DomainNet feature distribution skew | accuracy (Avg): 33.95 -> 35.99 | Helped |
| Use FedLMG with specific-teacher distillation instead of FGL | FedLMG on Unique NICO++ · Unique NICO++ feature distribution skew | accuracy (Avg): 74.11 -> 75.96 | Helped |
| Use FedLMG with specific-teacher distillation instead of FGL | FedLMG on Common NICO++ · Common NICO++ feature distribution skew | accuracy (Avg): 54.79 -> 58.16 | Helped |
| Add BN loss and CLS loss (ablation) | FedLMG generation on DomainNet · DomainNet feature distribution skew | accuracy (Avg): 28.60 -> 34.28 | Helped |
| Add BN loss alongside CLS loss (ablation) | FedLMG generation on DomainNet · DomainNet feature distribution skew | accuracy (Avg): 31.11 -> 34.28 | Helped |
| Add CLS loss alongside BN loss (ablation) | FedLMG generation on DomainNet · DomainNet feature distribution skew | accuracy (Avg): 32.64 -> 34.28 | Helped |
| Add BN loss without CLS loss (ablation) | FedLMG generation on DomainNet · DomainNet feature distribution skew | accuracy (Avg): 28.60 -> 32.64 | Helped |
| Add CLS loss without BN loss (ablation) | FedLMG generation on DomainNet · DomainNet feature distribution skew | accuracy (Avg): 28.60 -> 31.11 | Helped |
Surprising findings
- The one-shot method outperforms the centralized training ceiling on OpenImage and DomainNet in several settings.
- Prompts-only generation without client models performs far worse than with client guidance, indicating client models are essential.
Limitations
- Relies on a pretrained Stable Diffusion model on the server and its safety checker.
- Ceiling baseline and some comparisons are on feature-skew settings only; label-skew results are discussed but limited.
- Theoretical bound relies on an assumption of bounded KL divergence between distributions that is hard to verify.
Experimental setup
Datasets: OpenImage DomainNet NICO++
Benchmarks: OpenImage DomainNet Common NICO++ Unique NICO++
Metrics: classification accuracy communication cost (M parameters) client FLOPs (G)
Baselines: FedAvg FedDF FedProx FedDyn Prompts Only FedDISC FGL Ceiling (centralized training)
Builds on: classifier guidance FedDF FedDISC FGL DENSE
Closest papers by meaning
- Exploring One-Shot Semi-supervised Federated Learning with Pre-trained Diffusion Models (AAAI 2024 · similarity 0.91)
- One-shot Federated Learning via Synthetic Distiller-Distillate Communication (NeurIPS 2024 · similarity 0.90)
- FedBiP: Heterogeneous One-Shot Federated Learning with Personalized Latent Diffusion Models (CVPR 2025 · similarity 0.90)
- Revisiting Ensembling in One-Shot Federated Learning (NeurIPS 2024 · similarity 0.90)
- Federated Model Heterogeneous Matryoshka Representation Learning (NeurIPS 2024 · similarity 0.89)
- FedQS: Optimizing Gradient and Model Aggregation for Semi-Asynchronous Federated Learning (NeurIPS 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.