Research map / Federated learning and privacy / Communication-efficient FL
FedPop: Federated Population-based Hyperparameter Tuning
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 FedPop, a population-based evolutionary "tuning-while-training" method that jointly tunes server-side and client-side federated hyperparameters, outperforming RS, SHA and FedEx on FL benchmarks.
Problem
Hyperparameter tuning in federated learning is costly and unreliable under limited client compute and distributed, Non-IID validation data; prior FL tuning methods handle only client-side HPs or use fixed search spaces and retraining.
Approach
Runs multiple federated tuning processes in parallel and applies online evolutionary updates (exploit top members, perturb their HPs and copy weights) at two levels: FedPop-G across configurations (server and client HPs) and FedPop-L within each round across active clients' client-side HP vectors, avoiding retraining after tuning.
Key results
- On CIFAR-10 with SHA wrapper, FedPop reaches 81.47% (IID) and 76.42% (NIID Dir1.0) global accuracy, versus 78.57% and 70.37% without tuning-method baseline improvements.
- On CIFAR-10 NIID Dir0.5 with RS wrapper, FedPop improves over FedEx by about 3.85% on average across NIID settings (RS) and 4.79% (SHA).
- On cross-silo PACS, OfficeHome, DomainNet, FedPop with SHA achieves 75.17%, 45.71%, 73.59% global accuracy, surpassing FedEx by up to 2.72%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedPop (population-based tuning) instead of RS as tuning algorithm, RS wrapper | CIFAR-10 global model · CIFAR-10 IID / NIID Dir1.0 / NIID Dir0.5 | accuracy (global): +2.14 IID, +4.78 Dir1.0, +4.13 Dir0.5 over RS (computed from table; reported text gives 2.14 IID) | Helped |
| Use FedPop instead of FedEx (SHA wrapper) | CIFAR-10 global model · CIFAR-10 IID / NIID Dir1.0 / NIID Dir0.5 | accuracy (global): +1.64 IID, +4.40 Dir1.0, +5.19 Dir0.5 | Helped |
| Use FedPop instead of FedEx (SHA wrapper) | Shakespeare · Shakespeare IID / NIID | accuracy (global): +1.69 IID, +1.81 NIID | Helped |
| Use FedPop instead of FedEx (SHA wrapper) | FEMNIST · FEMNIST IID / NIID | accuracy (global): +1.12 IID vs FedEx; +0.50 NIID vs FedEx; but IID vs SHA only +0.52 | Mixed |
| Use FedPop instead of FedEx on real-world cross-silo benchmarks (SHA wrapper) | PACS, OfficeHome, DomainNet (ResNet-18) · PACS / OfficeHome / DomainNet | accuracy (global): +1.70 PACS, +2.72 OfficeHome, +1.91 DomainNet | Helped |
| Add FedPop-G (inter-configuration tuning) to RS (ablation) | CIFAR-10 RS wrapper · CIFAR-10 IID / NIID Dir1.0 / NIID Dir0.5 | accuracy (global): +1.57 IID, +3.34 Dir1.0, +2.75 Dir0.5 | Helped |
| Add FedPop-L (intra-configuration tuning) to RS (ablation) | CIFAR-10 RS wrapper · CIFAR-10 IID / NIID Dir1.0 / NIID Dir0.5 | accuracy (global): +0.99 IID, +4.03 Dir1.0, +1.36 Dir0.5 | Helped |
| Combine FedPop-G and FedPop-L (vs either alone) (ablation) | CIFAR-10 RS wrapper · CIFAR-10 IID / NIID Dir1.0 / NIID Dir0.5 | accuracy (global): +0.57 IID, +1.44 Dir1.0, +1.38 Dir0.5 over FedPop-G alone (best single component) | Helped |
| Use FedPop instead of FedEx/RS with FedProx Fed-Opt | CIFAR-10 with FedProx · CIFAR-10 IID / NIID Dir1.0 / NIID Dir0.5 | accuracy (global): FedPop vs FedEx: +2.15 IID, +1.96 Dir1.0, +2.55 Dir0.5 | Helped |
| Use FedPop instead of FedEx/RS with SCAFFOLD Fed-Opt | CIFAR-10 with SCAFFOLD · CIFAR-10 IID / NIID Dir1.0 / NIID Dir0.5 | accuracy (global): FedPop vs FedEx: +0.61 IID, +2.64 Dir1.0, +2.33 Dir0.5 | Helped |
Surprising findings
- Directly sampling client HP-vectors from the full prior distribution led to unstable model performance, so client HPs are sampled within a small ball around a base vector.
Limitations
- Relies on a fixed tuning budget and predefined HP search spaces and perturbation intensity schedules.
- Reported gains in some settings are within overlapping standard deviations, e.g., FEMNIST IID with RS.
Experimental setup
Datasets: CIFAR-10 FEMNIST Shakespeare PACS OfficeHome DomainNet ImageNet-1K
Benchmarks: CIFAR-10 FEMNIST Shakespeare PACS OfficeHome DomainNet ImageNet-1K
Metrics: testing accuracy (global and locally finetuned) standard deviation over 5 trials
Baselines: Random Search Successive Halving FedEx FLoRA
Builds on: Population-based Training FedAvg FedEx Successive Halving FedProx SCAFFOLD
Closest papers by meaning
- Adaptive Federated Learning with Auto-Tuned Clients (ICLR 2024 · similarity 0.88)
- Probabilistic Federated Prompt-Tuning with Non-IID and Imbalanced Data (NeurIPS 2024 · similarity 0.87)
- FedCal: Achieving Local and Global Calibration in Federated Learning via Aggregated Parameterized Scaler (ICML 2024 · similarity 0.86)
- FedQS: Optimizing Gradient and Model Aggregation for Semi-Asynchronous Federated Learning (NeurIPS 2025 · similarity 0.86)
- Federated Full-Parameter Tuning of Billion-Sized Language Models with Communication Cost under 18 Kilobytes (ICML 2024 · similarity 0.86)
- Federated Adaptive Prompt Tuning for Multi-Domain Collaborative Learning (AAAI 2024 · 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.