Research map / Federated learning and privacy / Communication-efficient FL
FedCompass: Efficient Cross-Silo Federated Learning on Heterogeneous Client Devices Using a Computing Power-Aware Scheduler
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 FedCompass, a semi-asynchronous cross-silo FL algorithm whose computing-power-aware scheduler assigns per-client local step counts so client updates arrive in near-simultaneous groups for aggregation.
Problem
Cross-silo federated learning with heterogeneous client compute and non-IID data: synchronous FL wastes time on stragglers, while asynchronous FL suffers from stale local models and client drift.
Approach
A Compass scheduler estimates each client's speed from past response times and assigns varying local step counts within [Qmin, Qmax] so clients are grouped to arrive nearly simultaneously; group aggregation reduces staleness while the overall process stays asynchronous.
Key results
- FedCompass reaches target accuracy fastest in all tested MNIST and CIFAR-10 settings; e.g., FedAvg needs up to 11.5x and FedAT up to 5.35x-ish relative time in some settings, while FedAsync/FedBuff often fail to reach target accuracy.
- FedAsync and FedBuff fail to reach target accuracy in many low-heterogeneity settings (marked '-') due to client drift from staleness weighting.
- Theory: convergence to a first-order stationary point with rate O(m/sqrt(T)) under bounded heterogeneity and staleness assumptions.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FedCompass instead of FedBuff | federated learning · MNIST Class Partition (90%), m=5, homo | relative wall-clock time to target accuracy: 1.30× -> 1.00× | Helped |
| use FedCompass instead of FedBuff | federated learning · MNIST Dirichlet Partition (90%), m=5, homo | relative wall-clock time to target accuracy: 1.23× -> 1.00× | Helped |
| use FedCompass instead of FedAT | federated learning · CIFAR-10 Class Partition (60%), m=5, homo | relative wall-clock time to target accuracy: 1.05× -> 1.00× | Helped |
| use FedCompass instead of FedAT | federated learning · Fed-IXI (0.98), homo | relative wall-clock time to target accuracy: 1.07× -> 1.00× | Helped |
| use FedCompass instead of FedAsync | federated learning · Fed-ISIC2019 (50%), normal | relative wall-clock time to target accuracy: 1.75× -> 1.00× | Helped |
| add server-side momentum (ablation) | FedCompass · MNIST and CIFAR-10 partitions across client counts and heterogeneity settings | relative wall-clock time to target accuracy: 0.72×; 1.19× | Mixed |
| add normalized averaging (ablation) | FedCompass · MNIST and CIFAR-10 partitions across client counts and heterogeneity settings | relative wall-clock time to target accuracy: 0.77×; 1.55× | Mixed |
| vary 1/Q (Qmin/Qmax) from 0.05 to 0.8 (ablation) | FedCompass · MNIST and CIFAR-10 Dirichlet partitions | relative wall-clock time to target accuracy: 0.89/1.27×; 1.58/3.44× | Mixed |
| set Qmin=Qmax (1/Q=1), removing flexible local-step assignment (ablation) | FedCompass · MNIST and CIFAR-10 Dirichlet partitions | relative wall-clock time to target accuracy: 1.78×/-; -/- | Hurt |
Surprising findings
- FedAsync and FedBuff sometimes fail to reach target accuracy at all in low-heterogeneity settings, attributed to unnecessary staleness penalization causing client drift.
- Server momentum or normalized averaging applied within FedCompass sometimes further speeds convergence, while FedCompass+M is slower on Fed-ISIC2019.
Limitations
- Client heterogeneity and speed changes are simulated with random distributions rather than real hardware.
- Convergence guarantees rely on bounded gradient, bounded variance, and bounded heterogeneity/staleness assumptions.
- Not suited to cross-device FL where client participation is intermittent.
Experimental setup
Datasets: MNIST CIFAR-10 Fed-IXI Fed-ISIC2019
Benchmarks: MNIST class partition MNIST dual Dirichlet partition CIFAR-10 class partition CIFAR-10 dual Dirichlet partition Fed-IXI Fed-ISIC2019
Metrics: relative wall-clock time to target validation accuracy validation accuracy
Baselines: FedAvg FedAvgM FedAsync FedBuff FedAT
Builds on: FedAvg FedAsync FedBuff FedAT FedAvgM
Closest papers by meaning
- FedQS: Optimizing Gradient and Model Aggregation for Semi-Asynchronous Federated Learning (NeurIPS 2025 · similarity 0.88)
- Efficient Federated Learning against Heterogeneous and Non-stationary Client Unavailability (NeurIPS 2024 · similarity 0.88)
- FedEL: Federated Elastic Learning for Heterogeneous Devices (NeurIPS 2025 · similarity 0.88)
- FedLPS: Heterogeneous Federated Learning for Multiple Tasks with Local Parameter Sharing (AAAI 2024 · similarity 0.88)
- TRAIL: Trust-Aware Client Scheduling for Semi-Decentralized Federated Learning (AAAI 2025 · similarity 0.88)
- FedLMT: Tackling System Heterogeneity of Federated Learning via Low-Rank Model Training with Theoretical Guarantees (ICML 2024 · similarity 0.88)
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.