Research map / Federated learning and privacy / Personalized and heterogeneous FL
LoGoFair: Post-Processing for Local and Global Fairness 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 LoGoFair, a model-agnostic federated post-processing framework that calibrates a pretrained FL classifier toward the Bayes-optimal classifier under joint local and global DP/EO fairness constraints.
Problem
Fairness in federated learning is usually addressed for either local (per-client) or global (pooled) group fairness, but under statistical heterogeneity they can conflict, and retraining federated models for fairness is costly and privacy-sensitive.
Approach
Derives a closed-form Bayes-optimal fair classifier under composite linear disparity constraints for both local and global fairness (a threshold on a calibrated score shifted by Lagrange multipliers), then learns those multipliers with a federated bilevel procedure using log-exp smoothing and projected gradient descent on client validation data.
Key results
- On Adult (alpha=0.5), LoGoFair_l&g reaches 0.8214 accuracy with local DP 0.0489 and global DP 0.0204, versus FedAvg 0.8381 / 0.1922 / 0.1759.
- LoGoFair_g achieves global DP 0.0183 on Adult at alpha=0.5, against 0.0767 for FedFB and 0.1479 for FCFL.
- Accuracy loss relative to FedAvg is small (roughly 1-2 points) while both local and global DP gaps drop substantially.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply both local and global fairness constraints rather than FedAvg | LoGoFair l&g · Adult (α=0.5) | local DP disparity: 0.1922 -> 0.0489 | Helped |
| apply both local and global fairness constraints rather than FedAvg | LoGoFair l&g · Adult (α=0.5) | global DP disparity: 0.1759 -> 0.0204 | Helped |
| apply both local and global fairness constraints rather than FedAvg | LoGoFair l&g · ENEM (α=0.5) | local DP disparity: 0.2030 -> 0.0281 | Helped |
| apply both local and global fairness constraints rather than FedAvg | LoGoFair l&g · ENEM (α=0.5) | global DP disparity: 0.1953 -> 0.0151 | Helped |
| apply both local and global fairness constraints rather than FedAvg | LoGoFair l&g · CelebA (α=0.5) | local DP disparity: 0.1308 -> 0.0312 | Helped |
| apply both local and global fairness constraints rather than FedAvg | LoGoFair l&g · CelebA (α=0.5) | global DP disparity: 0.1441 -> 0.0295 | Helped |
| use local fairness constraint only instead of both local and global constraints (ablation) | LoGoFair · Adult (α=0.5) | local DP disparity: 0.0489 -> 0.0367 | Helped |
| use local fairness constraint only instead of both local and global constraints (ablation) | LoGoFair · Adult (α=0.5) | global DP disparity: 0.0204 -> 0.0468 | Hurt |
| use global fairness constraint only instead of both local and global constraints (ablation) | LoGoFair · ENEM (α=0.5) | global DP disparity: 0.0151 -> 0.0141 | Helped |
| use global fairness constraint only instead of both local and global constraints (ablation) | LoGoFair · ENEM (α=0.5) | local DP disparity: 0.0281 -> 0.0574 | Hurt |
Surprising findings
- Local fairness does not imply global fairness and vice versa under statistical heterogeneity; methods targeting one notion degrade on the other as heterogeneity grows.
Limitations
- Relies on validation data at clients and estimated Bayes score; estimation error grows as client count increases (local fairness slightly worse on Adult and CelebA).
- Results reported for a fixed delta setting and 5 clients in main experiments; the method's trade-off depends on chosen fairness relaxation levels.
Experimental setup
Datasets: Adult ENEM CelebA
Benchmarks: Adult ENEM CelebA
Metrics: accuracy local DP global DP local EO global EO
Baselines: FedAvg FedFB FairFed FCFL
Builds on: FedAvg model calibration composite linear disparity
Closest papers by meaning
- FedFACT: A Provable Framework for Controllable Group-Fairness Calibration in Federated Learning (NeurIPS 2025 · similarity 0.93)
- Demystifying Local & Global Fairness Trade-offs in Federated Learning Using Partial Information Decomposition (ICLR 2024 · similarity 0.90)
- Multi-Dimensional Fair Federated Learning (AAAI 2024 · similarity 0.89)
- FairTrade: Achieving Pareto-Optimal Trade-Offs between Balanced Accuracy and Fairness in Federated Learning (AAAI 2024 · similarity 0.89)
- FedLF: Layer-Wise Fair Federated Learning (AAAI 2024 · similarity 0.87)
- Does Egalitarian Fairness Lead to Instability? The Fairness Bounds in Stable Federated Learning Under Altruistic Behaviors (NeurIPS 2024 · 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.