Research map / Federated learning and privacy / Personalized and heterogeneous FL
Multi-Dimensional Fair 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 mFairFL, a federated learning method that enforces group fairness via Lagrangian constraints and resolves client gradient conflicts with loss-ordered gradient projection before aggregation, evaluated on Adult, COMPAS and Bank.
Problem
Training fair federated models that satisfy group fairness across sensitive groups while also ensuring client fairness, given heterogeneous client data and gradient conflicts that disadvantage some clients.
Approach
Relax the group-fairness-constrained ERM with Lagrange multipliers, then at the server detect conflicting client gradients by cosine similarity with an EMA similarity goal and iteratively project gradients (ordered by client loss) to mitigate conflicts before averaging.
Key results
- On Adult under high heterogeneity, mFairFL reaches accuracy .792 with DP violation .012 and client-fairness violation .036, versus FedFB at .779/.014 and FedMGDA+ at .837/.238.
- On Bank, mFairFL achieves accuracy .844 with client-fairness violation .028, compared with CenFL's .866 accuracy and .001 DP violation.
- Ablation shows loss-ordered projection beats random and reversed orders (Adult CF violation .036 vs .049 and .064).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use loss-based projection-target ordering instead of random ordering (ablation) | mFairFL · Adult, COMPAS, and Bank | client fairness violation: Adult: .049 -> .036; Compas: .044 -> .022; Bank: .047 -> .028 | Helped |
| Use loss-based projection-target ordering instead of reverse ordering (ablation) | mFairFL · Adult, COMPAS, and Bank | client fairness violation: Adult: .064 -> .036; Compas: .049 -> .022; Bank: .077 -> .028 | Helped |
| Use loss-based projection-target ordering instead of random ordering (ablation) | mFairFL · Adult, COMPAS, and Bank | accuracy: Adult: .774 -> .792; Compas: .577 -> .596; Bank: .837 -> .844 | Helped |
| Use loss-based projection-target ordering instead of reverse ordering (ablation) | mFairFL · Adult, COMPAS, and Bank | accuracy: Adult: .768 -> .792; Compas: .580 -> .596; Bank: .822 -> .844 | Helped |
| Use mFairFL rather than FedMGDA+ | federated learning · Adult, COMPAS, and Bank (high heterogeneity) | accuracy: Adult: .837 -> .792; Compas: .635 -> .596; Bank: .874 -> .844 | Mixed |
| Use mFairFL rather than FedMGDA+ | federated learning · Adult, COMPAS, and Bank (high heterogeneity) | client fairness violation: Adult: .063 -> .036; Compas: .044 -> .022; Bank: .065 -> .028 | Helped |
| Use mFairFL rather than FedAvg-F | federated learning · Adult, COMPAS, and Bank (high heterogeneity) | accuracy: Adult: .706 -> .792; Compas: .558 -> .596; Bank: .828 -> .844 | Helped |
| Use mFairFL rather than FedAvg-F | federated learning · Adult, COMPAS, and Bank (high heterogeneity) | DP violation: Adult: .224 -> .012; Compas: .059 -> .005; Bank: .033 -> .005 | Helped |
Surprising findings
- FedMGDA+ achieves the highest accuracy in several settings but has substantially worse group fairness violations.
- FedAvg-f sometimes shows lower accuracy than the purely local IndFair baseline.
Limitations
- Experiments use only three small tabular datasets and five simulated clients with a shallow MLP.
- Theoretical convergence result assumes L-smooth objectives and is stated loosely; the paper's theorems are not tied to the practical non-convex setting.
- Hyperparameters beta and delta are tuned via cross-validation, and the method adds server-side pairwise gradient computation cost.
Experimental setup
Datasets: Adult COMPAS Bank
Benchmarks: Adult COMPAS Bank
Metrics: accuracy demographic parity violation equalized odds violation accuracy parity violation client fairness violation
Baselines: IndFair FedAvg-f CenFair FedFB FPFL FairFed FedMGDA+ q-FFL DRFL Ditto
Builds on: FedAvg FedMGDA+ q-FFL PCGrad
Closest papers by meaning
- FedFACT: A Provable Framework for Controllable Group-Fairness Calibration in Federated Learning (NeurIPS 2025 · similarity 0.89)
- LoGoFair: Post-Processing for Local and Global Fairness in Federated Learning (AAAI 2025 · similarity 0.89)
- FedLF: Layer-Wise Fair Federated Learning (AAAI 2024 · similarity 0.89)
- Fair Federated Learning under Domain Skew with Local Consistency and Domain Diversity (CVPR 2024 · similarity 0.89)
- FairTrade: Achieving Pareto-Optimal Trade-Offs between Balanced Accuracy and Fairness in Federated Learning (AAAI 2024 · similarity 0.88)
- Demystifying Local & Global Fairness Trade-offs in Federated Learning Using Partial Information Decomposition (ICLR 2024 · 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.