Research map / Federated learning and privacy / Personalized and heterogeneous FL
FedAWA: Adaptive Optimization of Aggregation Weights in Federated Learning Using Client Vectors
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 FedAWA, which adaptively optimizes server-side federated aggregation weights using client vectors (local model minus global model) without any proxy dataset.
Problem
Data heterogeneity in federated learning makes dataset-size-based FedAvg aggregation weights suboptimal, and existing adaptive weighting methods either are fixed or need a proxy dataset that raises privacy concerns.
Approach
Replace task vectors from model merging with client vectors (local model minus global model) and optimize aggregation weights so that clients whose update directions align with the merged global vector get higher weight, with a cosine-based regularizer keeping the global model close to the previous round; no extra data is needed.
Key results
- FedAWA achieves 49.48% average top-1 accuracy across CIFAR-10, CIFAR-100, and TinyImageNet settings, vs 47.88% for FedAvg and 48.26% for FedLAW (which uses a proxy dataset).
- FedAWA-L (layer-wise) reaches 49.98% average accuracy, the best overall in Table 1.
- On CIFAR-10 with alpha=0.1, FedAWA reaches 63.55% vs 61.04% for FedAvg; FedAWA aggregation takes 0.82 s vs 10.11 s for FedLAW and 2.52 s for L-DAWA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| adaptively optimize aggregation weights using client vectors | FedAWA · CIFAR-10, CIFAR-100, Tiny-ImageNet (α=0.1, 0.5, 100) | Top-1 test accuracy | Mixed |
| use layer-wise rather than shared aggregation weights (ablation) | FedAWA · CIFAR-10, CIFAR-100, Tiny-ImageNet | Top-1 test accuracy: 65.13 vs 63.55; 42.52 vs 41.89; 33.42 vs 33.07 | Helped |
| remove the global-alignment regularization term (ablation) | FedAWA · CIFAR-10, NIID (α=0.1) | Top-1 test accuracy: 63.05 vs 63.55 | Hurt |
| use Euclidean distance instead of 1 - cosine similarity for regularization (ablation) | FedAWA · CIFAR-10, NIID (α=0.1) | Top-1 test accuracy: 63.53 vs 63.55 | No clear effect |
| use cosine similarity directly as aggregation weights instead of optimizing them (ablation) | FedAWA · CIFAR-10 and CIFAR-100 | Top-1 test accuracy: 62.14 vs 63.55; 42.31 vs 42.84 | Hurt |
| use layer-wise aggregation weights (ablation) | FedAWA · CIFAR-10, ResNet20 | aggregation execution time: 15.21 vs 0.82 | Hurt |
Surprising findings
- Directly using cosine similarity as aggregation weights (FedAWA-COS) is less effective than optimizing the weights, and FedLAW with a proxy dataset is outperformed by FedAWA on average.
Limitations
- Applicable only when client model architectures are identical; heterogeneous architectures are not handled.
- FedAWA-L layer-wise variant incurs higher computational overhead (15.21 s aggregation time).
Experimental setup
Datasets: CIFAR-10 CIFAR-100 TinyImageNet
Benchmarks: CIFAR-10 CIFAR-100 TinyImageNet
Metrics: top-1 test accuracy aggregation execution time similarity to ideal aggregation weights
Baselines: FedAvg FedProx FedDyn FedAdam FedDisco FedLAW L-DAWA
Builds on: FedAvg Task Arithmetic FedLAW L-DAWA FedDisco
Closest papers by meaning
- FedLWS: Federated Learning with Adaptive Layer-wise Weight Shrinking (ICLR 2025 · similarity 0.91)
- FedAA: A Reinforcement Learning Perspective on Adaptive Aggregation for Fair and Robust Federated Learning (AAAI 2025 · similarity 0.90)
- An Aggregation-Free Federated Learning for Tackling Data Heterogeneity (CVPR 2024 · similarity 0.90)
- Principled Federated Domain Adaptation: Gradient Projection and Auto-Weighting (ICLR 2024 · similarity 0.89)
- FedQS: Optimizing Gradient and Model Aggregation for Semi-Asynchronous Federated Learning (NeurIPS 2025 · similarity 0.89)
- TurboSVM-FL: Boosting Federated Learning through SVM Aggregation for Lazy Clients (AAAI 2024 · 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.