Research map / Federated learning and privacy / Communication-efficient FL
FedASMU: Efficient Asynchronous Federated Learning with Dynamic Staleness-Aware Model Update
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 FedASMU, an asynchronous federated learning framework with dynamic staleness-aware server aggregation (polynomial staleness weights with learned control parameters) and device-side adaptive mixing of fresh global models chosen by an RL/Q-learning time-slot selector.
Problem
Federated learning under statistical heterogeneity (non-IID data) and system heterogeneity (device stragglers); asynchronous FL suffers accuracy loss from stale uploaded models and non-IID data.
Approach
Dynamically adjust the aggregation weight of each uploaded model at the server using a staleness-dependent polynomial whose control parameters are updated by gradient descent on the global loss, and let devices mid-training fetch and blend a fresh global model at an RL-selected epoch, with the blend weight also adapted by gradient steps.
Key results
- FedASMU reports accuracy 0.60% to 23.90% higher than baselines and training time 3.54% to 97.98% faster across six models and five datasets.
- Example: on LeNet CIFAR-10 FedASMU reaches 0.486 accuracy in 8800 time units versus 0.431 and 125514 for FedAvg.
- Ablation: removing dynamic model aggregation lowers accuracy by 1.38% to 4.32%; removing fresh global model aggregation lowers accuracy by 0.65% to 3.04%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedASMU instead of synchronous or asynchronous baseline approaches | Federated learning · Five public datasets and six models | accuracy: 0.60\% to 23.90\% higher | Helped |
| Use FedASMU instead of baseline approaches | Federated learning · Five public datasets and six models | training speed: 3.54\% to 97.98\% faster | Helped |
| Add dynamic model aggregation on the server (ablation) | FedASMU · CIFAR-10 | accuracy: 1.38\% to 4.32\% | Helped |
| Add dynamic model aggregation on the server (ablation) | FedASMU-0 · CIFAR-10 | accuracy: 0.65\% to 3.04\% | Helped |
| Add fresh global model aggregation on devices (ablation) | FedASMU-FA · CIFAR-10 | training time to target accuracy: 44.77\% to 73.96\% | Helped |
| Add fresh global model aggregation on devices (ablation) | FedASMU-FA · CIFAR-10 | accuracy: 1.75\% to 4.71\% | Helped |
| Add fresh global model aggregation on devices (ablation) | FedASMU-DA · CIFAR-10 | accuracy: 1.04\% to 3.41\% | Helped |
| Add fresh global model aggregation on devices (ablation) | FedASMU-DA · CIFAR-10 | training speed: 15.71\% to 19.54\% faster | Helped |
| Use asynchronous FedASMU-0 rather than FedAvg (ablation) | Federated learning · CIFAR-10 | accuracy: 0.08\% to 0.34\% | Helped |
| Use asynchronous FedASMU-0 rather than FedAvg (ablation) | Federated learning · CIFAR-10 | training speed: 67.84\% to 82.26\% faster | Helped |
Surprising findings
- Asynchronous FedAsync-style variant (FedASMU-0) gives slightly higher accuracy than FedAvg while being much faster.
Limitations
- Only one fresh global model is fetched per local training round to limit communication overhead, and the extra global-model transfer reduces benefit under limited bandwidth.
- Convergence analysis relies on strong convexity, L-smoothness and bounded gradient/variance assumptions that are strong for deep networks.
Experimental setup
Datasets: Fashion-MNIST CIFAR-10 CIFAR-100 IMDb Tiny-ImageNet
Benchmarks: Fashion-MNIST CIFAR-10 CIFAR-100 IMDb Tiny-ImageNet
Metrics: convergence accuracy training time to target accuracy
Baselines: FedAvg FedProx MOON FedDyn FedAsync PORT ASO-Fed FedBuff FedSA
Builds on: FedAsync FedAvg Q-learning REINFORCE
Closest papers by meaning
- Efficient Asynchronous Federated Learning with Prospective Momentum Aggregation and Fine-Grained Correction (AAAI 2024 · similarity 0.91)
- Tackling Intertwined Data and Device Heterogeneities in Federated Learning with Unlimited Staleness (AAAI 2025 · similarity 0.91)
- FedSMU: Communication-Efficient and Generalization-Enhanced Federated Learning through Symbolic Model Updates (ICML 2025 · similarity 0.90)
- FedQS: Optimizing Gradient and Model Aggregation for Semi-Asynchronous Federated Learning (NeurIPS 2025 · similarity 0.90)
- Reviving Stale Updates: Data-Free Knowledge Distillation for Asynchronous Federated Learning (NeurIPS 2026 · similarity 0.90)
- Momentum-Driven Adaptivity: Towards Tuning-Free Asynchronous Federated Learning (ICML 2025 · similarity 0.90)
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.