Research map / Federated learning and privacy / Communication-efficient FL
Fair Federated Learning via the Proportional Veto Core
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.
Adapts the proportional veto core from social choice to federated learning, proves PVC-stable models exist under Lebesgue-measurable utilities, and proposes Rank-Core-Fed to find an approximately PVC-stable model over a sampled representative set.
Problem
Core-stability fairness guarantees in federated learning require concave agent utilities, which fail for non-concave deep network losses, so utility-based fairness guarantees become impractical.
Approach
Measure model quality by ordinal rank rather than cardinal utility and adapt the proportional veto core from social choice theory to FL, yielding rankwise proportionality, Pareto efficiency, and existence under only Lebesgue-measurability of utilities.
Key results
- On CIFAR-10 with the low-heterogeneity instance, Rank-Core-Fed reaches beta = 0.95 versus 0.88 for FedAvg and 0.89 for CoreFed, with comparable social welfare (37.41 vs 37.29).
- On MNIST with the high-heterogeneity instance, Rank-Core-Fed achieves beta = 0.89 vs 0.86 for FedAvg and 0.87 for CoreFed, with welfare 9.23 comparable to baselines.
- Theorem 1 proves the epsilon-PVC is non-empty for any epsilon in (0, 1/n) whenever the feasible set and utilities are Lebesgue-measurable, with no concavity needed.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Rank-Core-Fed instead of the strongest baseline (q-FFL) | Federated learning · MNIST Ihigh | β: 0.87 -> 0.89 | Helped |
| Use Rank-Core-Fed instead of the strongest baseline (q-FFL) | Federated learning · MNIST Ihigh | sum of utilities: 9.26 -> 9.23 | No clear effect |
| Use Rank-Core-Fed instead of the strongest baseline (CoreFed) | Federated learning · CIFAR-10 Ihigh | β: 0.81 -> 0.85 | Helped |
| Use Rank-Core-Fed instead of the strongest baseline (q-FFL) | Federated learning · CIFAR-10 Ihigh | sum of utilities: 36.75 -> 36.81 | Helped |
| Use Rank-Core-Fed instead of the strongest baseline (CoreFed) | Federated learning · MNIST Imid | β: 0.89 -> 0.92 | Helped |
| Use Rank-Core-Fed instead of the strongest baseline (CoreFed) | Federated learning · MNIST Imid | sum of utilities: 9.56 -> 9.63 | Helped |
| Use Rank-Core-Fed instead of the strongest baseline (q-FFL) | Federated learning · CIFAR-10 Imid | β: 0.89 -> 0.91 | Helped |
| Use Rank-Core-Fed instead of the strongest baseline (CoreFed) | Federated learning · CIFAR-10 Imid | sum of utilities: 37.23 -> 37.19 | Hurt |
| Use Rank-Core-Fed instead of the strongest baseline (CoreFed) | Federated learning · MNIST Ilow | β: 0.93 -> 0.97 | Helped |
| Use Rank-Core-Fed instead of the strongest baseline (q-FFL) | Federated learning · MNIST Ilow | sum of utilities: 9.65 -> 9.70 | Helped |
| Use Rank-Core-Fed instead of the strongest baseline (q-FFL) | Federated learning · CIFAR-10 Ilow | β: 0.91 -> 0.95 | Helped |
| Use Rank-Core-Fed instead of the strongest baseline (FedAvg) | Federated learning · CIFAR-10 Ilow | sum of utilities: 37.29 -> 37.41 | Helped |
Surprising findings
- Concavity-based core stability is unattainable even for approximate proportionality with smooth non-concave utilities (Example 1), so no utility-based guarantee can hold in general.
Limitations
- Computing the full PVC over all models is intractable for general utilities; Rank-Core-Fed optimizes over a sampled representative set only.
- Efficient algorithms for computing PVC-stable models with continuous model sets and general utilities are unknown.
- Experiments are limited to small CNN and VGG11 models on rotated MNIST and CIFAR-10 with synthetic heterogeneity.
Experimental setup
Datasets: rotated MNIST CIFAR-10
Benchmarks: rotated MNIST CIFAR-10
Metrics: beta rankwise proportionality (minimum agent rank of final m utilitarian social welfare (sum of agent utilities)
Baselines: FedAvg CoreFed q-FFL PropFair TERM
Builds on: CoreFed FedAvg proportional veto core q-FFL
Closest papers by meaning
- You Get What You Give: Reciprocally Fair Federated Learning (ICML 2025 · similarity 0.88)
- Pursuing Overall Welfare in Federated Learning through Sequential Decision Making (ICML 2024 · similarity 0.88)
- Does Egalitarian Fairness Lead to Instability? The Fairness Bounds in Stable Federated Learning Under Altruistic Behaviors (NeurIPS 2024 · similarity 0.87)
- FairTrade: Achieving Pareto-Optimal Trade-Offs between Balanced Accuracy and Fairness in Federated Learning (AAAI 2024 · similarity 0.87)
- Fair Federated Learning under Domain Skew with Local Consistency and Domain Diversity (CVPR 2024 · similarity 0.86)
- FedFACT: A Provable Framework for Controllable Group-Fairness Calibration in Federated Learning (NeurIPS 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.