Research map / Federated learning and privacy / Personalized and heterogeneous FL
Fair Federated Learning under Domain Skew with Local Consistency and Domain Diversity
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 FedHEAL, a plug-in federated aggregation method that discards parameter updates with low directional consistency and reweights clients by distance-variance minimization to achieve performance fairness under domain skew.
Problem
Performance fairness in federated learning under domain skew, where parameter update conflicts and aggregation weighting bias cause uneven accuracy across domains/clients.
Approach
Observes Parameter Update Consistency (parameters repeatedly moving in the same direction across rounds) and uses it to keep only consistent, important parameter updates in aggregation (FPHL); client weights are updated via momentum toward larger-distance clients to minimize variance of client-global distances (FAEL).
Key results
- On Digits, FedAvg+FedHEAL improves average accuracy from 76.00 to 78.09 and reduces STD from 23.82 to 22.08.
- On Office-Caltech, FedAvg+FedHEAL reaches 64.85 average accuracy with STD 3.80, versus 58.35 and 11.13 for FedAvg.
- FedProto+FedHEAL achieves 74.30 average accuracy on Office-Caltech with STD 6.44, versus 59.18 and 12.04 without FedHEAL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedHEAL instead of FedFV | ResNet-10 federated learning · Digits | average accuracy: 76.97 -> 78.09 | Helped |
| Use FedHEAL instead of FedFV | ResNet-10 federated learning · Digits | standard deviation of accuracy: 23.17 -> 22.08 | Helped |
| Use FedHEAL instead of FedFV | ResNet-10 federated learning · Office-Caltech | average accuracy: 62.41 -> 64.85 | Helped |
| Use FedHEAL instead of FedFV | ResNet-10 federated learning · Office-Caltech | standard deviation of accuracy: 6.72 -> 3.80 | Helped |
| Add FPHL, omitting FAEL (ablation) | FedHEAL · Digits | average accuracy: 76.00 -> 77.05 | Helped |
| Add FAEL, omitting FPHL (ablation) | FedHEAL · Digits | average accuracy: 76.00 -> 77.20 | Helped |
| Add both FPHL and FAEL (ablation) | FedHEAL · Digits | average accuracy: 76.00 -> 78.09 | Helped |
| Add FPHL, omitting FAEL (ablation) | FedHEAL · Office-Caltech | average accuracy: 58.35 -> 64.73 | Helped |
| Add FAEL, omitting FPHL (ablation) | FedHEAL · Office-Caltech | average accuracy: 58.35 -> 60.19 | Helped |
| Add both FPHL and FAEL (ablation) | FedHEAL · Office-Caltech | average accuracy: 58.35 -> 64.85 | Helped |
Surprising findings
- Converged global models still show strong parameter update consistency over the last 10 rounds, suggesting they have not adapted to specific domains.
Limitations
- Performance is sensitive to hyperparameters tau and beta, and may be unstable if they are poorly chosen.
- Assumes all clients share the same network architecture; may fail when architectures differ.
Experimental setup
Datasets: Digits (MNIST, USPS, SVHN, SYN) Office-Caltech
Benchmarks: Digits Office-Caltech
Metrics: average accuracy standard deviation of accuracy across domains
Baselines: FedAvg FedProx Scaffold MOON FedDyn FedProc FedProto AFL q-FFL FedFV Ditto
Builds on: FedAvg q-FFL AFL FedFV
Closest papers by meaning
- FairTrade: Achieving Pareto-Optimal Trade-Offs between Balanced Accuracy and Fairness in Federated Learning (AAAI 2024 · similarity 0.90)
- FedLF: Layer-Wise Fair Federated Learning (AAAI 2024 · similarity 0.90)
- Demystifying Local & Global Fairness Trade-offs in Federated Learning Using Partial Information Decomposition (ICLR 2024 · similarity 0.89)
- Multi-Dimensional Fair Federated Learning (AAAI 2024 · similarity 0.89)
- FedFACT: A Provable Framework for Controllable Group-Fairness Calibration in Federated Learning (NeurIPS 2025 · similarity 0.88)
- FedAA: A Reinforcement Learning Perspective on Adaptive Aggregation for Fair and Robust 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.