Research map / Federated learning and privacy / Communication-efficient FL
DualGFL: Federated Learning with a Dual-Level Coalition-Auction Game
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 DualGFL, a federated learning framework combining a lower-level hedonic coalition-formation game (with a Pareto-optimal partitioning algorithm) and an upper-level multi-attribute procurement auction with greedy coalition selection, improving both server and client utility on FMNIST, EMNIST and CIFAR10.
Problem
Existing game-theoretical federated learning methods use one-level games (purely cooperative or competitive) and usually improve either client utility or server utility, not both, limiting realism for large-scale deployments.
Approach
A dual-level game: clients form coalitions (each with one edge server) via a hedonic game using an auction-aware utility that accounts for expected auction competitiveness, and coalitions bid in a multi-attribute procurement auction with resource (bandwidth) constraints, with the server selecting winners by a score-to-resource greedy rule.
Key results
- On FMNIST (0.1), DualGFL reaches total score 516.29 (50.58x over FedAvgHed benchmark) and average client utility 447.59 (4.08x over FedAvg).
- On EMNIST (0.1), DualGFL achieves average coalition quality 12293.74 (17.64x) and client utility 55.38 (12.51x over FedAvg).
- On CIFAR10 (0.1), DualGFL improves test accuracy to 75.17% versus 73.50% for FedAvg and 73.93% for FedAvgAuc.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DualGFL instead of FedAvgHed | Federated learning · FMNIST (0.6) | Total Score: 13.97 -> 544.87 | Helped |
| Use DualGFL instead of FedAvgAuc | Federated learning · FMNIST (0.6) | Avg. Client Quality: 1581.94 -> 1271.56 | Hurt |
| Use DualGFL instead of DualGFLStat | Federated learning · FMNIST (0.6) | Avg. Client Utility: 272.71 -> 483.06 | Helped |
| Use DualGFL instead of FedAvgHed | Federated learning · FMNIST (0.1) | Test Accuracy: 86.46% -> 88.43% | Helped |
| Use DualGFL instead of FedAvgHed | Federated learning · EMNIST (0.1) | Avg. Client Utility: 28.01 -> 55.38 | Helped |
| Use DualGFL instead of FedAvgAuc | Federated learning · CIFAR10 (0.1) | Test Accuracy: 73.93% -> 75.17% | Helped |
| Replace random winning-coalition selection with score-based selection (ablation) | DualGFL · FMNIST (0.1) | Total Score: 331.71 -> 516.29 | Helped |
| Increase maximum coalition size |S|_max (ablation) | DualGFL · FMNIST (0.1) | Avg. Coalition Quality | Helped |
| Increase maximum coalition size |S|_max (ablation) | DualGFL · FMNIST (0.1) | Avg. Client Quality | Hurt |
| Increase maximum coalition size |S|_max (ablation) | DualGFL · FMNIST (0.1) | Avg. Client Utility | Helped |
| Set maximum coalition size |S|_max to 8 (ablation) | DualGFL · FMNIST (0.1) | Total Score | Helped |
Surprising findings
- FedAvgAuc attains the highest average client quality but low coalition quality because it lacks coalition formation.
- The total score peaks at max coalition size 8 rather than at the largest tested size, attributed to balancing client movement and avoiding a Matthew effect.
Limitations
- Experiments are simulated on image classification with a shallow CNN only.
- Equilibrium bid analysis relies on assumptions such as i.i.d. private cost factors and a known distribution F.
- Client utility and payoff depend on the central server's quality metric (data size), which is a simplified quality measure.
Experimental setup
Datasets: FMNIST EMNIST CIFAR10
Benchmarks: FMNIST EMNIST CIFAR10
Metrics: test accuracy total score average client quality average coalition quality average client utility
Baselines: FedAvg FedAvgAuc FedAvgHed DualGFLStat
Builds on: FedAvg hierarchical federated learning Flower
Closest papers by meaning
- Multi-Session Budget Optimization for Forward Auction-based Federated Learning (ICML 2025 · similarity 0.87)
- Reputation-aware Revenue Allocation for Auction-based Federated Learning (AAAI 2025 · similarity 0.87)
- Learning Coalition Structures with Games (AAAI 2024 · similarity 0.87)
- Multiplayer Federated Learning: Reaching Equilibrium with Less Communication (NeurIPS 2025 · similarity 0.86)
- Pursuing Overall Welfare in Federated Learning through Sequential Decision Making (ICML 2024 · similarity 0.86)
- Convergence Analysis of Split Federated Learning on Heterogeneous Data (NeurIPS 2024 · similarity 0.85)
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.