Research map / Federated learning and privacy / Communication-efficient FL
Communication-Efficient Federated Learning with Accelerated Client Gradient
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 FedACG, a federated learning method that broadcasts a lookahead global model (global model plus scaled global momentum) and regularizes local updates toward it, improving convergence and communication efficiency under low client participation.
Problem
Federated learning suffers from slow, unstable convergence and client drift under non-i.i.d. client data, worsened by low client participation rates.
Approach
The server broadcasts a single lookahead model (previous global model plus scaled global momentum), so clients start local training from a point aligned with the global update direction, and a proximal term keeps local models close to this accelerated point; no extra communication, server computation, or client memory is needed.
Key results
- On CIFAR-100 with 100 clients and 5% participation (Dirichlet 0.3), FedACG reaches 62.51% at 1000 rounds versus 58.37% for MOON and 53.29% for FedAvgM.
- With 500 clients and 2% participation on Tiny-ImageNet, FedACG reaches the 30% target in 447 rounds, while FedAvg does not reach it within 1000 rounds.
- On FEMNIST, FedACG reaches 80.61% accuracy at 500 rounds, above FedDC's 80.11%, while sending 56.8% fewer parameters than FedDC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedACG’s momentum-integrated client initialization and local regularization | Federated learning · CIFAR-10, 100 clients, 5% participation, 1000R | accuracy: 85.48 -> 89.10 | Helped |
| Use FedACG’s momentum-integrated client initialization and local regularization | Federated learning · CIFAR-100, 100 clients, 5% participation, 1000R | accuracy: 59.01 -> 62.51 | Helped |
| Use FedACG’s momentum-integrated client initialization and local regularization | Federated learning · Tiny-ImageNet, 100 clients, 5% participation, 1000R | accuracy: 45.51 -> 46.31 | Helped |
| Use FedACG’s momentum-integrated client initialization and local regularization | Federated learning · CIFAR-10, 500 clients, 2% participation, 1000R | accuracy: 83.49 -> 82.80 | Hurt |
| Use FedACG’s momentum-integrated client initialization and local regularization | Federated learning · CIFAR-100, 500 clients, 2% participation, 1000R | accuracy: 45.93 -> 48.40 | Helped |
| Use FedACG’s momentum-integrated client initialization and local regularization | Federated learning · Tiny-ImageNet, 500 clients, 2% participation, 1000R | accuracy: 33.67 -> 38.48 | Helped |
| Use FedACG’s momentum-integrated client initialization and local regularization | Federated learning · CIFAR-10, 100 clients, 2% participation, 1000R | accuracy: 82.86 -> 84.73 | Helped |
| Use FedACG’s momentum-integrated client initialization and local regularization | Federated learning · CIFAR-10, 500 clients, 1% participation, 1000R | accuracy: 75.06 -> 76.45 | Helped |
| Use FedACG’s momentum-integrated client initialization and local regularization | Federated learning · CIFAR-100, 100 clients, 2% participation, 1000R | accuracy: 53.47 -> 56.89 | Helped |
| Use FedACG’s momentum-integrated client initialization and local regularization | Federated learning · CIFAR-100, 500 clients, 1% participation, 1000R | accuracy: 39.64 -> 45.18 | Helped |
| Add server update with momentum (ablation) | FedACG · CIFAR-10 and CIFAR-100, 500 clients, 2% participation, 1000R | accuracy: 71.45 -> 82.80; 38.11 -> 48.40 | Helped |
| Add accelerated client gradient (ablation) | FedACG · CIFAR-10 and CIFAR-100, 500 clients, 2% participation, 1000R | accuracy: 70.75 -> 82.80; 36.16 -> 48.40 | Helped |
Surprising findings
- Using the local regularization term alone does not reliably help on CIFAR-10 and CIFAR-100, but combined with accelerated client gradient it gives a larger gain.
- FedCM performs poorly on the Shakespeare next-word task even after extensive hyperparameter tuning.
Limitations
- Gains are reported mainly on image and small-scale federated benchmarks with ResNet-18 and LSTM backbones.
- Convergence analysis relies on assumptions (bounded gradient dissimilarity) that the authors note are standard for momentum methods.
- Sensitivity to hyperparameters lambda and beta is shown only on CIFAR-10.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 Tiny-ImageNet FEMNIST CelebA Shakespeare
Benchmarks: CIFAR-10 CIFAR-100 Tiny-ImageNet FEMNIST CelebA Shakespeare
Metrics: test accuracy communication rounds to target accuracy weight divergence layer-wise CKA
Baselines: FedAvg FedProx FedAvgM FedADAM FedDyn MOON FedCM FedMLB FedLC FedNTD FedDC FedDecorr
Builds on: FedAvg FedAvgM FedProx FedCM FedDC
Closest papers by meaning
- FedSMU: Communication-Efficient and Generalization-Enhanced Federated Learning through Symbolic Model Updates (ICML 2025 · similarity 0.90)
- Low Precision Local Training is Enough for Federated Learning (NeurIPS 2024 · similarity 0.90)
- FedQS: Optimizing Gradient and Model Aggregation for Semi-Asynchronous Federated Learning (NeurIPS 2025 · similarity 0.89)
- Layer-wise Update Aggregation with Recycling for Communication-Efficient Federated Learning (NeurIPS 2025 · similarity 0.89)
- Efficient Federated Learning against Heterogeneous and Non-stationary Client Unavailability (NeurIPS 2024 · similarity 0.89)
- FedAdaVR: Adaptive Variance Reduction for Robust Federated Learning under Limited Client Participation (NeurIPS 2026 · similarity 0.89)
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.