Research map / Federated learning and privacy / Communication-efficient FL
Efficient Asynchronous Federated Learning with Prospective Momentum Aggregation and Fine-Grained Correction
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 FedAC, an asynchronous federated learning framework combining temporal-gradient client weighting, Nadam-style prospective momentum at the server, and fine-grained gradient correction at clients, evaluated on CIFAR-10, EMNIST-L and Shakespeare.
Problem
Asynchronous federated learning is inefficient because data heterogeneity and device heterogeneity cause poor client training quality, stale updates, and slow server convergence.
Approach
Weights each client by the cosine similarity between its pseudo-gradient and the global pseudo-gradient over the staleness interval, applies a Nadam-style momentum with adaptive learning rate on buffered asynchronous aggregation, and corrects every local SGD step with a SCAFFOLD-like global-minus-local control term computed from the client weights.
Key results
- On CIFAR-10 with target accuracy 0.7, FedAC needs 2800 client trips versus 10000 for FedAsync (3.6x) and 4180 for FedProx (1.5x).
- On Shakespeare with target accuracy 0.4, FedAC needs 1080 client trips versus 2320 for FedAvg (2.1x) and 6290 for FedAsync (5.8x).
- At 3000 client trips on CIFAR-10 with alpha=0.1, FedAC reaches 70.6% accuracy versus 66.8% for FedAvg and 63.1% for FedBuff.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedAC instead of FedAvg | CIFAR-10 · CIFAR-10, α=0.1 | accuracy: 66.8 -> 70.6 | Helped |
| Use FedAC instead of FedAvg | CIFAR-10 · CIFAR-10, α=1 | accuracy: 76.1 -> 79.6 | Helped |
| Use FedAC instead of FedAdam | CIFAR-10 · CIFAR-10, α=10 | accuracy: 78.8 -> 80.1 | Helped |
| Use FedAC instead of FedAdam | CIFAR-10 · CIFAR-10, IID | accuracy: 79.1 -> 80.3 | Helped |
| Add client weight evaluation based on temporal gradient (ablation) | FedBuff-based framework · CIFAR-10 | accuracy | Helped |
| Add adaptive server update with prospective weighted momentum (ablation) | Basic-WTG · CIFAR-10 | convergence speed | Helped |
| Add adaptive server update with prospective weighted momentum without fine-grained client correction (ablation) | Basic-WTG · CIFAR-10 | accuracy | Hurt |
| Add fine-grained gradient correction during client updates (ablation) | FedAC · CIFAR-10 | accuracy | Helped |
Limitations
- Experiments limited to small models (MLP, small CNN, character LSTM) and three datasets.
- Communication cost of the extra correction term (client sends delta c_i and gradient) is not analyzed; the paper notes the ideal update requires communication with all clients.
- FedAdam required hyperparameters borrowed from its original paper because it performed poorly under the shared settings.
Experimental setup
Datasets: CIFAR-10 EMNIST-L Shakespeare
Benchmarks: CIFAR-10 EMNIST-L Shakespeare
Metrics: accuracy number of client trips to target accuracy accuracy at fixed client trips
Baselines: FedAvg FedAdam FedProx FedAsync FedBuff
Builds on: FedBuff FedAdam SCAFFOLD FedAsync Nadam
Closest papers by meaning
- Momentum-Driven Adaptivity: Towards Tuning-Free Asynchronous Federated Learning (ICML 2025 · similarity 0.93)
- FedASMU: Efficient Asynchronous Federated Learning with Dynamic Staleness-Aware Model Update (AAAI 2024 · similarity 0.91)
- Reviving Stale Updates: Data-Free Knowledge Distillation for Asynchronous Federated Learning (NeurIPS 2026 · similarity 0.91)
- FedQS: Optimizing Gradient and Model Aggregation for Semi-Asynchronous Federated Learning (NeurIPS 2025 · similarity 0.90)
- Communication-Efficient Federated Learning with Accelerated Client Gradient (CVPR 2024 · similarity 0.89)
- GAS: Generative Activation-Aided Asynchronous Split Federated Learning (AAAI 2025 · 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.