Research map / Federated learning and privacy / Personalized and heterogeneous FL
Enhancing One-Shot Federated Learning Through Data and Ensemble Co-Boosting
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 Co-Boosting, a one-shot federated learning method that mutually boosts synthetic hard-sample data and ensemble client weights, then distills the server model, without modifying local training or transmitting extra data.
Problem
One-shot federated learning (OFL) aggregates client models into a server model in a single round, but server performance depends on both synthetic data quality and ensemble quality, and existing methods improve only one of them or require local training changes or extra transmissions.
Approach
Generate hard synthetic samples adversarially from the current ensemble (hard-sample loss plus adversarial KL loss and on-the-fly perturbation), use these samples to learn ensemble mixing weights over client logits via sign-gradient updates, and iterate so data and ensemble co-improve before distilling the server model.
Key results
- On CIFAR-10 with Dir(0.05), Co-Boosting reaches 47.20% server accuracy vs 38.37% for the best baseline DENSE (+8.83%).
- On SVHN with Dir(0.05), Co-Boosting reaches 65.40% server accuracy, outperforming DENSE's 60.24%.
- The Co-Boosting ensemble improves over the averaged FedENS ensemble, e.g., 48.75% vs 42.34% on CIFAR-10 with Dir(0.05).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Co-Boosting instead of the strongest listed baseline, FedDF | one-shot federated learning server model · MNIST, Dir(0.1) | test accuracy: 87.91$\pm$0.92 -> 94.44$\pm$1.02 | Helped |
| use Co-Boosting instead of the strongest listed baseline, DENSE | one-shot federated learning server model · FMNIST, Dir(0.1) | test accuracy: 69.43$\pm$1.94 -> 74.86$\pm$1.70 | Helped |
| use Co-Boosting instead of the strongest listed baseline, FedDF | one-shot federated learning server model · SVHN, Dir(0.1) | test accuracy: 68.98$\pm$0.63 -> 72.88$\pm$1.19 | Helped |
| use Co-Boosting instead of the strongest listed baseline, DENSE | one-shot federated learning server model · CIFAR-10, Dir(0.1) | test accuracy: 47.80$\pm$1.21 -> 57.09$\pm$0.94 | Helped |
| use Co-Boosting instead of the strongest listed baseline, DENSE | one-shot federated learning server model · CIFAR-100, Dir(0.1) | test accuracy: 22.21$\pm$1.41 -> 23.59$\pm$1.27 | Helped |
| add hard-sample generation loss (GHS) (ablation) | Co-Boosting · SVHN, Dir(0.05), 10 clients | test accuracy: 58.46 -> 61.18 | Helped |
| add on-the-fly diverse hard samples (DHS) (ablation) | Co-Boosting · SVHN, Dir(0.05), 10 clients | test accuracy: 58.46 -> 61.38 | Helped |
| add ensemble enhancement through reweighting (EE) (ablation) | Co-Boosting · SVHN, Dir(0.05), 10 clients | test accuracy: 58.46 -> 62.67 | Helped |
| add DHS and EE, without GHS (ablation) | Co-Boosting · SVHN, Dir(0.05), 10 clients | test accuracy: 58.46 -> 63.42 | Helped |
| add GHS and DHS, without EE (ablation) | Co-Boosting · SVHN, Dir(0.05), 10 clients | test accuracy: 58.46 -> 62.46 | Helped |
| add GHS and EE, without DHS (ablation) | Co-Boosting · SVHN, Dir(0.05), 10 clients | test accuracy: 58.46 -> 64.40 | Helped |
| add GHS, DHS, and EE together (ablation) | Co-Boosting · SVHN, Dir(0.05), 10 clients | test accuracy: 58.46 -> 65.40 | Helped |
Limitations
- Ensemble mixing weights are learned only on synthetic samples, leaving a gap versus weights learned on real data.
- Evaluated only on small image classification datasets with small CNN/ResNet-style models.
- Relies on the availability of well-pretrained client models and their logits; no comparison with methods requiring auxiliary public data.
Experimental setup
Datasets: MNIST FMNIST SVHN CIFAR-10 CIFAR-100 Tiny-ImageNet MNIST-M PACS
Benchmarks: MNIST FMNIST SVHN CIFAR-10 CIFAR-100
Metrics: test accuracy
Baselines: FedAvg FedDF F-ADI F-DAFL DENSE FedENS DW-FedENS
Builds on: DENSE FedDF DAFL ADI FedAvg
Closest papers by meaning
- FuseFL: One-Shot Federated Learning through the Lens of Causality with Progressive Model Fusion (NeurIPS 2024 · similarity 0.87)
- One-shot Federated Learning via Synthetic Distiller-Distillate Communication (NeurIPS 2024 · similarity 0.87)
- FedLPA: One-shot Federated Learning with Layer-Wise Posterior Aggregation (NeurIPS 2024 · similarity 0.86)
- Revisiting Ensembling in One-Shot Federated Learning (NeurIPS 2024 · similarity 0.86)
- Does One-shot Give the Best Shot? Mitigating Model Inconsistency in One-shot Federated Learning (ICML 2025 · similarity 0.86)
- FedBEns: One-Shot Federated Learning based on Bayesian Ensemble (ICML 2025 · similarity 0.86)
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.