Research map / Federated learning and privacy / Personalized and heterogeneous FL
Adapted Weighted Aggregation in 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 FedAW, a federated learning aggregation method that weights client models by a bias-based data-quality score from biased and debiased local models, evaluated on Colored MNIST.
Problem
Federated learning under non-iid data where clients hold datasets with different amounts of bias-conflicting samples, which degrades the global model when FedAvg averages all clients equally.
Approach
Scores each client's dataset quality via the ratio of cross-entropy losses of a biased model and a debiased model (in the style of LfF), then aggregates client models with weights proportional to these scores (linear v1 or softmax v2) instead of uniform FedAvg weights.
Key results
- On Colored MNIST with the LfF debiasing setup, FedAw v1 reaches 0.86 accuracy versus 0.80 for FedAvg (LF row of Table 1).
- FedAw v1 gets 0.65 vs 0.63 for FedAvg on the Vanilla setting (Table 1).
- Disparate impact for FedAw v1 under Vanilla is 1.6 vs 2.1 for FedAvg (Table 2).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace FedAvg aggregation with FedAW v1 linear score-based weighting | Federated model aggregation · CMNIST fair test set (Vanilla) | accuracy: 0.63 -> 0.65 | Helped |
| Replace FedAvg aggregation with FedAW v1 linear score-based weighting | Federated model aggregation · CMNIST fair test set (LF) | accuracy: 0.80 -> 0.86 | Helped |
| Replace FedAvg aggregation with FedAW v1 linear score-based weighting | Federated model aggregation · CMNIST fair test set (BiasAdv) | accuracy: 0.78 -> 0.80 | Helped |
| Replace FedAvg aggregation with FedAW v2 softmax score-based weighting | Federated model aggregation · CMNIST fair test set (Vanilla) | accuracy: 0.63 -> 0.64 | Helped |
| Replace FedAvg aggregation with FedAW v2 softmax score-based weighting | Federated model aggregation · CMNIST fair test set (LF) | accuracy: 0.80 -> 0.83 | Helped |
| Replace FedAvg aggregation with FedAW v2 softmax score-based weighting | Federated model aggregation · CMNIST fair test set (BiasAdv) | accuracy: 0.78 -> 0.80 | Helped |
| Use linear rather than FedAvg aggregation | Federated model aggregation · CMNIST fair test set (Vanilla) | disparate impact: 2.1 -> 1.6 | Helped |
| Use softmax rather than FedAvg aggregation | Federated model aggregation · CMNIST fair test set (Vanilla) | disparate impact: 2.1 -> 1.7 | Helped |
| Use linear rather than FedAvg aggregation | Federated model aggregation · CMNIST fair test set (LF) | disparate impact: 1.0 -> 1.0 | No clear effect |
| Use softmax rather than FedAvg aggregation | Federated model aggregation · CMNIST fair test set (LF) | disparate impact: 1.0 -> 0.96 | Helped |
| Use linear rather than FedAvg aggregation | Federated model aggregation · CMNIST fair test set (BiasAdv) | disparate impact: 1.1 -> 1.0 | Helped |
| Use softmax rather than FedAvg aggregation | Federated model aggregation · CMNIST fair test set (BiasAdv) | disparate impact: 1.1 -> 1.6 | Hurt |
Surprising findings
- The BiasAdv row shows FedAw v2 tying FedAvg-level performance with v1 only slightly higher, and the paper's conclusion mentions FedBN efficacy though FedBN is not evaluated in the experiments.
Limitations
- Evaluated only on Colored MNIST, a single synthetic benchmark with 10 clients.
- Scheduling of when to start reweighting (e.g., 2000 epochs for CMNIST) is set case by case without a general rule.
- Weighting strategy choice is left to users; no principled selection among v1, v2, or polynomial weights.
Experimental setup
Datasets: Colored MNIST
Benchmarks: Colored MNIST
Metrics: accuracy disparate impact
Baselines: FedAvg q-FFL
Builds on: FedAvg LfF q-FFL
Closest papers by meaning
- FedAWA: Adaptive Optimization of Aggregation Weights in Federated Learning Using Client Vectors (CVPR 2025 · similarity 0.86)
- Principled Federated Domain Adaptation: Gradient Projection and Auto-Weighting (ICLR 2024 · similarity 0.84)
- An Aggregation-Free Federated Learning for Tackling Data Heterogeneity (CVPR 2024 · similarity 0.84)
- A Lightweight Method for Tackling Unknown Participation Statistics in Federated Averaging (ICLR 2024 · similarity 0.83)
- FedAA: A Reinforcement Learning Perspective on Adaptive Aggregation for Fair and Robust Federated Learning (AAAI 2025 · similarity 0.83)
- FedECADO: A Dynamical System Model of Federated Learning (ICML 2025 · similarity 0.83)
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.