Research map / Federated learning and privacy / Personalized and heterogeneous FL
Federated Oriented Learning: A Practical One-Shot Personalized Federated Learning Framework
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 Federated Oriented Learning (FOL), a one-shot personalized federated learning framework for satellite-like settings where clients exchange models once with neighbors, using fine-tuning, alignment-aware pruning, ensemble weighting, and distillation.
Problem
Personalized federated learning requires many communication rounds, which is infeasible in settings with limited contact windows such as LEO satellite constellations; one-shot personalized learning with only neighbor model exchanges is needed.
Approach
Each client collects neighbor models, fine-tunes and structurally prunes them with an alignment regularizer toward its own model, post-fine-tunes them, selects top-K by validation accuracy (cosine tie-break), forms an optimized weighted logit ensemble, and distills it into a regularized personalized student model that is carried to the next round.
Key results
- On Wildfire at Dirichlet psi=0.1, FOL exceeds the best baseline by up to 39.24% (satellite #48).
- At psi=0.7 on Hurricane satellite #44, FOL surpasses the best baseline by 21.77%.
- On CIFAR-10, CIFAR-100, and SVHN, FOL beats the best baseline by up to 3.72%, 5.97%, and 9.27%, respectively.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply FOL instead of Co-Boosting | personalized federated learning · Wildfire, ψ=0.7, satellite #28 | test accuracy: 89.19 ± 1.13 -> 95.75 ± 1.39 | Helped |
| apply FOL instead of Co-Boosting | personalized federated learning · Hurricane, ψ=0.7, satellite #44 | test accuracy: 74.04 ± 1.54 -> 95.81 ± 1.88 | Helped |
| apply FOL instead of Co-Boosting | personalized federated learning · Wildfire, ψ=0.1, satellite #48 | test accuracy: 54.51 ± 1.85 -> 93.75 ± 1.40 | Helped |
| apply FOL instead of DENSE | personalized federated learning · Wildfire, ψ=0.3, satellite #43 | test accuracy: 78.63 ± 1.98 -> 92.37 ± 1.27 | Helped |
| apply FOL instead of Co-Boosting | personalized federated learning · CIFAR-10, ψ=0.7, satellite #9 | test accuracy: 63.11 ± 1.98 -> 66.83 ± 0.82 | Helped |
| apply FOL instead of Co-Boosting | personalized federated learning · CIFAR-100, ψ=0.7, satellite #14 | test accuracy: 33.45 ± 2.12 -> 39.42 ± 1.03 | Helped |
| apply FOL instead of Co-Boosting | personalized federated learning · SVHN, ψ=0.5, satellite #21 | test accuracy: 73.58 ± 1.48 -> 82.85 ± 1.18 | Helped |
| add fine-tuning, pruning, and post-pruning fine-tuning to the ensemble variant (ablation) | FOL-A · CIFAR-10, ψ=0.7, satellite #9 | test accuracy: 63.01 ± 0.71 -> 71.02 ± 0.73 | Helped |
| add fine-tuning, pruning, and post-pruning fine-tuning to the ensemble variant (ablation) | FOL-A · CIFAR-100, ψ=0.7, satellite #14 | test accuracy: 32.05 ± 1.68 -> 49.24 ± 1.12 | Helped |
| add fine-tuning, pruning, and post-pruning fine-tuning to the ensemble variant (ablation) | FOL-A · SVHN, ψ=0.5, satellite #21 | test accuracy: 83.15 ± 1.33 -> 88.37 ± 0.92 | Helped |
| add fine-tuning, pruning, and post-pruning fine-tuning to the personalized variant (ablation) | FOL · CIFAR-10, ψ=0.7, satellite #9 | test accuracy: 61.49 ± 1.43 -> 66.83 ± 0.82 | Helped |
| add fine-tuning, pruning, and post-pruning fine-tuning to the personalized variant (ablation) | FOL · CIFAR-100, ψ=0.7, satellite #14 | test accuracy: 31.11 ± 2.12 -> 39.42 ± 1.03 | Helped |
Surprising findings
- Direct fine-tuning of a global one-shot model is argued to perform arbitrarily poorly for personalization, and FedAvg performs worse than local training in most reported settings.
Limitations
- Evaluation uses only a subset of clients (one to three satellites per configuration) due to compute constraints.
- Baselines DENSE and Co-Boosting are run with 5 clients, not the 70-client setting used for FOL.
- No privacy guarantees are provided; the authors note this as future work.
Experimental setup
Datasets: Wildfire Hurricane CIFAR-10 CIFAR-100 SVHN
Benchmarks: Wildfire Hurricane CIFAR-10 CIFAR-100 SVHN
Metrics: test accuracy
Baselines: FedAvg DENSE Co-Boosting Local training
Builds on: FedAvg DENSE Co-Boosting
Closest papers by meaning
- FedAS: Bridging Inconsistency in Personalized Federated Learning (CVPR 2024 · similarity 0.86)
- Adaptive Latent-Space Constraints in Personalized Federated Learning (NeurIPS 2025 · similarity 0.86)
- Decentralized Directed Collaboration for Personalized Federated Learning (CVPR 2024 · similarity 0.86)
- One-shot Federated Learning via Synthetic Distiller-Distillate Communication (NeurIPS 2024 · similarity 0.86)
- Understanding the Statistical Accuracy-Communication Trade-off in Personalized Federated Learning with Minimax Guarantees (ICML 2025 · similarity 0.86)
- FedSPU: Personalized Federated Learning for Resource-Constrained Devices with Stochastic Parameter Update (AAAI 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.