Research map / Federated learning and privacy / Personalized and heterogeneous FL
FedGCR: Achieving Performance and Fairness for Federated Learning with Distinct Client Types via Group Customization and Reweighting
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 FedGCR, a federated learning method for distinct client types that combines prompt-tuned ViT customization (FedGC) with group-wise reweighting aggregation (FedGR) using anonymized GMM clustering of client representations.
Problem
Federated learning where clients belong to distinct groups (client types, e.g., different data domains), where existing methods focus on individual-client performance and fairness and ignore group-level disparities and customization.
Approach
Uses type-specific prompts generated by GC-Net from frozen ViT features, added to global visual prompts, plus a fairness-aware aggregation that reweights clients by both their own loss and their anonymously inferred cluster's average loss, with a scheduled balance parameter.
Key results
- FedGCR reaches 82.74% average accuracy with sigma_type 12.26 on Digit-Five at DIF=10, versus 60.62% and 26.76 for FedAvg.
- On DomainNet at DIF=10, FedGCR reaches 81.88% accuracy and sigma_type 14.88, outperforming FedSAM at 81.04%.
- Clustering accuracy reaches 100% on Digit-Five when the cluster count is at least the number of client types (T' >= 5).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedGCR instead of FedSAM | Federated learning on Digit-Five · Digit-Five, DIF=10 | Avg accuracy: 64.06 (1.47) -> 82.74 (0.47) | Helped |
| Use FedGCR instead of AFL | Federated learning on Digit-Five · Digit-Five, DIF=10 | σtype: 23.58 (0.49) -> 12.26 (0.36) | Helped |
| Use FedGCR instead of AFL | Federated learning on Digit-Five · Digit-Five, DIF=10 | σclient: 18.70 (0.27) -> 8.52 (0.19) | Helped |
| Use FedGCR instead of FedSAM | Federated learning on DomainNet · DomainNet, DIF=10 | Avg accuracy: 81.04 (0.28) -> 81.88 (0.52) | Helped |
| Use FedGCR instead of AFL | Federated learning on DomainNet · DomainNet, DIF=10 | σtype: 16.83 (0.19) -> 14.88 (0.08) | Helped |
| Use FedGCR instead of AFL | Federated learning on DomainNet · DomainNet, DIF=10 | σclient: 11.09 (0.04) -> 9.07 (0.17) | Helped |
| Add FedGR to FedGC (ablation) | FedGCR on Digit-Five · Digit-Five, DIF=10 | Avg accuracy: 79.49 -> 82.74 | Helped |
| Add FedGR to FedGC (ablation) | FedGCR on Digit-Five · Digit-Five, DIF=10 | σdomain: 16.57 -> 12.26 | Helped |
| Add FedGR to FedGC (ablation) | FedGCR on Digit-Five · Digit-Five, DIF=10 | σclient: 11.61 -> 8.52 | Helped |
| Add FedGC to FedGR (ablation) | FedGCR on Digit-Five · Digit-Five, DIF=10 | Avg accuracy: 63.01 -> 82.74 | Helped |
| Add FedGC to FedGR (ablation) | FedGCR on Digit-Five · Digit-Five, DIF=10 | σdomain: 23.03 -> 12.26 | Helped |
| Add FedGC to FedGR (ablation) | FedGCR on Digit-Five · Digit-Five, DIF=10 | σclient: 17.75 -> 8.52 | Helped |
Surprising findings
- Setting the cluster count slightly above the true number of client types (T'=6) gave the best results on Digit-Five.
Limitations
- Client types are assumed static; dynamic or hybrid client types are not studied.
- Evaluation is limited to image classification with domain-based client types.
- Requires a predetermined cluster count T'.
Experimental setup
Datasets: Digit-Five DomainNet PACS
Benchmarks: Digit-Five DomainNet
Metrics: average accuracy sigma_type sigma_client
Baselines: FedAvg q-FFL AFL TERM Harmo-FL FedSAM FedMix
Builds on: Visual Prompt Tuning q-FFL FedAvg Gaussian Mixture Model clustering
Closest papers by meaning
- FedGPS: Statistical Rectification Against Data Heterogeneity in Federated Learning (NeurIPS 2025 · similarity 0.88)
- FedFACT: A Provable Framework for Controllable Group-Fairness Calibration in Federated Learning (NeurIPS 2025 · similarity 0.88)
- Benchmarking Algorithms for Federated Domain Generalization (ICLR 2024 · similarity 0.88)
- FedLF: Layer-Wise Fair Federated Learning (AAAI 2024 · similarity 0.87)
- Principled Federated Domain Adaptation: Gradient Projection and Auto-Weighting (ICLR 2024 · similarity 0.87)
- Interaction-Aware Gaussian Weighting for Clustered Federated Learning (ICML 2025 · similarity 0.86)
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.