Research map / Federated learning and privacy / Communication-efficient FL
FedQS: Optimizing Gradient and Model Aggregation for Semi-Asynchronous Federated Learning
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 FedQS, a semi-asynchronous federated learning framework that jointly optimizes gradient (FedSGD-style) and model (FedAvg-style) aggregation by classifying clients into four types by update speed and update similarity, with theoretical convergence bounds and experiments on CIFAR-10, Shakespeare, and UCI Adult.
Problem
Semi-asynchronous FL suffers a performance gap between gradient aggregation (fast but fluctuating) and model aggregation (stable but slow and less accurate), with no theory or unified method addressing both.
Approach
A divide-and-conquer scheme that classifies clients into four quadrants (fast/straggling x strongly/weakly biased) using local update speed and local-global gradient similarity, adapting learning rates, momentum, and server aggregation weights per type, usable with both gradient and model aggregation.
Key results
- FedQS-Avg reaches 68.88% average accuracy in the CV-based model-aggregation table, 1.71% above the best model-aggregation baseline M-step; FedQS-SGD reaches 72.06%, 2.51% above WKAFL.
- FedQS reduces runtime by 70.34% (Avg) and 70.91% (SGD) relative to synchronous FedAvg and FedSGD.
- Across dynamic scenarios, FedQS-SGD improves accuracy over FedSGD (e.g., 76.17% vs 68.29% at x=0.1 in Scenario 1).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedQS-Avg instead of M-step | SAFL model aggregation · CIFAR-10 x=0.5 | accuracy: 80.49 -> 80.26 | Hurt |
| Use FedQS-SGD instead of the strongest baseline | SAFL gradient aggregation · CIFAR-10 x=0.1 | accuracy: 65.71 -> 68.88 | Helped |
| Use FedQS-SGD instead of the strongest baseline | SAFL gradient aggregation · Shakespeare R=600 | accuracy: 51.62 -> 52.49 | Helped |
| Replace cosine similarity with another similarity measure (ablation) | FedQS · CIFAR-10 x=0.1, 0.5, 1 average | accuracy: 74.14; 75.69; 76.56 | Mixed |
| Add momentum to local training (ablation) | FedQS-Avg and FedQS-SGD · CIFAR-10 x=0.1, 0.5, 1 average | accuracy: 73.21 -> 74.14; 78.88 -> 80.59 | Helped |
| Add momentum to local training (ablation) | FedQS-Avg and FedQS-SGD · CIFAR-10 x=0.1, 0.5, 1 average | convergence speed (# epochs): 269 -> 251; 242 -> 230 | Helped |
| Add feedback mechanism (ablation) | FedQS-Avg · CIFAR-10 x=0.1, 0.5, 1 average | accuracy: 68.35 -> 74.14 | Helped |
| Add feedback mechanism (ablation) | FedQS-SGD · CIFAR-10 x=0.1, 0.5, 1 average | accuracy: 78.83 -> 80.59 | Helped |
Surprising findings
- Disabling the feedback mechanism in the model-aggregation variant reduces accuracy by 7.81%, more than in the gradient variant (2.18%).
Limitations
- Model-aggregation mode shows some oscillations.
- Introduces three extra hyperparameters (a, m0, k) that require tuning.
- Experiments limited to moderate-scale models; no differential privacy evaluation.
Experimental setup
Datasets: CIFAR-10 Shakespeare UCI Adult
Benchmarks: CIFAR-10 Shakespeare UCI Adult
Metrics: accuracy loss convergence speed (# epochs to target accuracy) runtime number of oscillations
Baselines: FedAvg SAFA FedAT M-step FedSGD FedBuff WKAFL FedAC DeFedAvg FADAS CA2FL
Builds on: FedAvg FedSGD FedBuff FedAT
Closest papers by meaning
- FedGPS: Statistical Rectification Against Data Heterogeneity in Federated Learning (NeurIPS 2025 · similarity 0.90)
- FedASMU: Efficient Asynchronous Federated Learning with Dynamic Staleness-Aware Model Update (AAAI 2024 · similarity 0.90)
- Convergence Analysis of Split Federated Learning on Heterogeneous Data (NeurIPS 2024 · similarity 0.90)
- Decentralized Sporadic Federated Learning: A Unified Algorithmic Framework with Convergence Guarantees (ICLR 2025 · similarity 0.90)
- TurboSVM-FL: Boosting Federated Learning through SVM Aggregation for Lazy Clients (AAAI 2024 · similarity 0.90)
- Efficient Asynchronous Federated Learning with Prospective Momentum Aggregation and Fine-Grained Correction (AAAI 2024 · 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.