Research map / Federated learning and privacy / Personalized and heterogeneous FL
FedFACT: A Provable Framework for Controllable Group-Fairness Calibration 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 FedFACT, a multi-class federated group-fairness calibration framework with in-processing (personalized cost-sensitive learning) and post-processing (bi-level optimization) methods that approach Bayes-optimal fair classifiers under global and local fairness constraints.
Problem
Achieving both global and local group fairness, in multi-class settings, with a controllable and near-optimal accuracy-fairness trade-off in federated learning, where fairness criteria are non-decomposable and non-differentiable.
Approach
Characterizes federated Bayes-optimal fair classifiers under global and local confusion-matrix constraints; reduces in-processing to personalized cost-sensitive learning with calibrated losses built from dual variables, and post-processing to a bi-level optimization over dual parameters using a plug-in estimate of the Bayes score.
Key results
- On Compas (gamma=0.5), FedFACT post-processing (g & l) reaches Acc 67.33 with global disparity 0.0139 and local 0.0641, versus FedAvg 69.73 / 0.2766 / 0.3590 and Cost 64.51 / 0.0585 / 0.0941.
- On Adult (gamma=0.5), FedFACT g-in achieves global DP disparity 0.0015, far below the best baseline Cost at 0.0262, with accuracy 82.05 vs 81.09.
- In multi-class experiments (CelebA, ENEM), in-processing gives better fairness calibration than post-processing, the reverse of the binary results.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedFACT post-processing with global and local fairness constraints instead of unconstrained FedAvg | Federated classification · Compas, γ=0.5 | D_global: 0.2766 -> 0.0139 | Helped |
| Use FedFACT post-processing with global and local fairness constraints instead of unconstrained FedAvg | Federated classification · Adult, γ=0.5 | D_local: 0.2310 -> 0.0274 | Helped |
| Use FedFACT post-processing with global and local fairness constraints instead of unconstrained FedAvg | Federated classification · CelebA, Hetero | D_global: 0.1360 -> 0.0088 | Helped |
| Use FedFACT post-processing with global and local fairness constraints instead of unconstrained FedAvg | Federated classification · ENEM, Hetero | D_local: 0.3068 -> 0.0131 | Helped |
| Relax the local fairness constraint from ξᶫ=0.00 to ξᶫ=0.04 at fixed ξᵍ=0.04 | FedFACT in-processing · Compas | Acc: 61.81 -> 62.39 | Helped |
| Relax the local fairness constraint from ξᶫ=0.00 to ξᶫ=0.04 at fixed ξᵍ=0.04 | FedFACT in-processing · Compas | D_local: 0.0836 -> 0.1105 | Hurt |
| Relax the global fairness constraint from ξᵍ=0.00 to ξᵍ=0.04 at fixed ξᶫ=0.00 | FedFACT in-processing · Compas | Acc: 61.17 -> 61.81 | Helped |
| Relax the global fairness constraint from ξᵍ=0.00 to ξᵍ=0.04 at fixed ξᶫ=0.00 | FedFACT in-processing · Compas | D_global: 0.0407 -> 0.0600 | Hurt |
| Use post-processing rather than in-processing | FedFACT with global and local constraints · Compas, γ=0.5 | Acc: 61.17 -> 67.33 | Helped |
| Use post-processing rather than in-processing | FedFACT with global and local constraints · Compas, γ=0.5 | D_local: 0.0732 -> 0.0641 | Hurt |
| Use post-processing rather than in-processing | FedFACT with global and local constraints · ENEM, Hetero | D_local: 0.0364 -> 0.0131 | Helped |
Surprising findings
- In multi-class settings post-processing underperforms in-processing, attributed to too few local samples for local fairness estimation.
- Trade-off controllability is only shown on Adult and Compas, not on all datasets.
Limitations
- Post-processing suffers from generalization error when local samples are scarce, particularly in multi-class settings.
- Combining in- and post-processing is theoretically unjustified, as stated by the authors.
- Assumes continuity of the Bayes score distribution (eta-continuity) for the post-processing characterization.
Experimental setup
Datasets: Compas Adult CelebA ENEM
Benchmarks: Compas Adult CelebA ENEM
Metrics: accuracy global fairness disparity local fairness disparity
Baselines: FedAvg FairFed FedFB FCFL praFFed Cost
Builds on: FedAvg Bayes-optimal fair classifier (centralized) Sion's minimax theorem
Closest papers by meaning
- LoGoFair: Post-Processing for Local and Global Fairness in Federated Learning (AAAI 2025 · similarity 0.93)
- Demystifying Local & Global Fairness Trade-offs in Federated Learning Using Partial Information Decomposition (ICLR 2024 · similarity 0.92)
- FairTrade: Achieving Pareto-Optimal Trade-Offs between Balanced Accuracy and Fairness in Federated Learning (AAAI 2024 · similarity 0.91)
- Multi-Dimensional Fair Federated Learning (AAAI 2024 · similarity 0.89)
- FedLF: Layer-Wise Fair Federated Learning (AAAI 2024 · similarity 0.89)
- FFB: A Fair Fairness Benchmark for In-Processing Group Fairness Methods (ICLR 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.