Research map / Federated learning and privacy / Communication-efficient FL
A Lightweight Method for Tackling Unknown Participation Statistics in Federated Averaging
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 FedAU, a lightweight FedAvg variant that adapts per-client aggregation weights online from each client's participation history (cutoff-interval averaging), converging to the original objective without storing per-client updates.
Problem
Federated averaging under unknown, uncontrollable, and heterogeneous client participation statistics, where non-optimal aggregation weights bias the solution toward frequently participating clients and existing fixes need N×d extra memory.
Approach
Estimate each client's optimal aggregation weight online as the average participation interval (with a cutoff K on interval length), so weights approximate 1/p_n without knowing p_n; the method needs only three scalar states per client and has a bias-variance tradeoff controlled by K.
Key results
- FedAU with K=50 reaches 86.0% train / 77.3% test on CIFAR-10 under Bernoulli participation, versus 83.5% / 74.1% for average-participating FedAvg.
- FedAU (K=50) outperforms FedVarp and MIFA, which use 250x memory, on most datasets and participation patterns, with CIFAR-10 test accuracy the exception where FedVarp is best.
- FedAU even slightly outperforms the known-participation-statistics baseline in several settings, which the authors attribute to using the empirical participation sample path.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use finite cutoff K=50 rather than K→∞ (ablation) | FedAU · SVHN test, Bernoulli | accuracy: 89.3{ $\pm$0.5} -> 89.6{ $\pm$0.4} | No clear effect |
| Use finite cutoff K=50 rather than K→∞ (ablation) | FedAU · CIFAR-10 test, Bernoulli | accuracy: 77.1{ $\pm$0.4} -> 77.3{ $\pm$0.3} | No clear effect |
| Use finite cutoff K=50 rather than K→∞ (ablation) | FedAU · CIFAR-100 test, Bernoulli | accuracy: 52.3{ $\pm$0.4} -> 52.1{ $\pm$0.6} | No clear effect |
| Use finite cutoff K=50 rather than K→∞ (ablation) | FedAU · CINIC-10 test, Bernoulli | accuracy: 61.5{ $\pm$0.4} -> 62.7{ $\pm$0.2} | Helped |
| Use finite cutoff K=50 rather than K→∞ (ablation) | FedAU · SVHN test, Markovian | accuracy: 89.3{ $\pm$0.4} -> 89.5{ $\pm$0.3} | No clear effect |
| Use finite cutoff K=50 rather than K→∞ (ablation) | FedAU · CIFAR-10 test, Markovian | accuracy: 77.1{ $\pm$0.3} -> 77.2{ $\pm$0.3} | No clear effect |
| Use finite cutoff K=50 rather than K→∞ (ablation) | FedAU · CIFAR-100 test, Markovian | accuracy: 51.8{ $\pm$0.3} -> 51.7{ $\pm$0.3} | No clear effect |
| Use finite cutoff K=50 rather than K→∞ (ablation) | FedAU · CINIC-10 test, Markovian | accuracy: 61.2{ $\pm$0.2} -> 62.3{ $\pm$0.2} | Helped |
| Use finite cutoff K=50 rather than K→∞ (ablation) | FedAU · SVHN test, Cyclic | accuracy: 88.7{ $\pm$0.6} -> 88.8{ $\pm$0.6} | No clear effect |
| Use finite cutoff K=50 rather than K→∞ (ablation) | FedAU · CIFAR-10 test, Cyclic | accuracy: 76.3{ $\pm$0.7} -> 76.6{ $\pm$0.4} | No clear effect |
| Use finite cutoff K=50 rather than K→∞ (ablation) | FedAU · CIFAR-100 test, Cyclic | accuracy: 50.6{ $\pm$0.3} -> 51.0{ $\pm$0.5} | No clear effect |
| Use finite cutoff K=50 rather than K→∞ (ablation) | FedAU · CINIC-10 test, Cyclic | accuracy: 60.0{ $\pm$0.8} -> 60.9{ $\pm$0.7} | Helped |
Surprising findings
- FedAU performs better than the idealized baseline that uses known participation probabilities, attributed to empirical sample-path deviations from p_n under low participation rates.
- Choosing a finite cutoff K=50 generally outperforms K→∞, and large K (>=500) degrades performance when some clients participate very rarely.
Limitations
- Theory assumes participation is independent across rounds; time-dependent (Markovian/cyclic) patterns are only validated empirically.
- Convergence bound requires either nearly optimal weights or a bounded global gradient assumption.
- Clients that participate only once cannot be handled without sharing information across clients.
Experimental setup
Datasets: SVHN CIFAR-10 CIFAR-100 CINIC-10
Benchmarks: SVHN CIFAR-10 CIFAR-100 CINIC-10
Metrics: accuracy
Baselines: FedAvg average participating FedAvg average all FedVarp MIFA known participation statistics
Builds on: FedAvg FedVarp MIFA
Closest papers by meaning
- Exact and Linear Convergence for Federated Learning under Arbitrary Client Participation is Attainable (NeurIPS 2025 · similarity 0.89)
- FedAWA: Adaptive Optimization of Aggregation Weights in Federated Learning Using Client Vectors (CVPR 2025 · similarity 0.88)
- Federated Learning under Periodic Client Participation and Heterogeneous Data: A New Communication-Efficient Algorithm and Analysis (NeurIPS 2024 · similarity 0.88)
- FedAdaVR: Adaptive Variance Reduction for Robust Federated Learning under Limited Client Participation (NeurIPS 2026 · similarity 0.88)
- Debiasing Federated Learning with Correlated Client Participation (ICLR 2025 · similarity 0.87)
- Understanding Server-Assisted Federated Learning in the Presence of Incomplete Client Participation (ICML 2024 · similarity 0.87)
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.