Research map / Federated learning and privacy / Personalized and heterogeneous FL
Free-Rider and Conflict Aware Collaboration Formation for Cross-Silo 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 FedEgoists, a coalition formation strategy for cross-silo federated learning that enforces no-free-rider and conflict-of-interest principles via clique and SCC-based clustering with merge operations, evaluated on CIFAR-10/100 and eICU against nine FL baselines.
Problem
Cross-silo federated learning among self-interested, competing organizations where free riding and conflicts of interest (contributing to competitors or their allies) must be avoided while forming collaboration groups.
Approach
Partition FL participants into coalitions by taking maximal cliques of the inverse competing graph (so members are mutually non-competing), then finding strongly connected components of the benefit graph within each clique, and merging coalitions via cycle, path and node conditions when this raises utility without violating the two principles.
Key results
- On CIFAR-10 with pathological partition at alpha=0.4, FedEgoists reaches 82.54±0.30 MTA versus 72.89 for the best baseline (FedOra).
- On CIFAR-100 with Dirichlet partition at alpha=0.2, FedEgoists reaches 32.86±1.53 MTA, above the best baseline FedOra at 27.24.
- On eICU, FedEgoists achieves average AUC 74.79 across ten hospitals, the highest among compared methods (best baseline FedOra 72.61).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedEgoists collaboration formation instead of the strongest baseline | Cross-silo FL on CIFAR-10 · CIFAR-10, pathological and Dirichlet distributions across α settings | MTA | Helped |
| Use FedEgoists collaboration formation instead of the strongest baseline | Cross-silo FL on CIFAR-100 · CIFAR-100, pathological and Dirichlet distributions across α settings | MTA | Mixed |
| Use FedEgoists collaboration formation instead of the strongest baseline | Cross-silo FL on eICU · eICU, Avg | AUC: 72.06 -> 74.79 | Helped |
Surprising findings
- Under the Dirichlet CIFAR-100 setting at alpha=0.05, the local baseline (30.31) outperforms FedEgoists (27.59), and FedEgoists does not win every setting on eICU (e.g., v0, v2, v5, v6 lower than some baselines).
Limitations
- Experiments use only n=10 FL participants and simulated competition graphs generated randomly.
- Benefit graph relies on a hypernetwork estimate of data complementarity, and the CS is assumed trustworthy.
- Worst-case complexity is exponential due to Bron-Kerbosch clique enumeration.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 eICU
Benchmarks: CIFAR-10 CIFAR-100 eICU
Metrics: mean test accuracy AUC worst-case performance difference
Baselines: Local FedAvg FedProx SCAFFOLD pFedMe pFedHN FedDisco pFedGraph FedOra
Builds on: Bron-Kerbosch algorithm Tarjan algorithm FedAvg hypernetwork benefit graph
Closest papers by meaning
- FedCompetitors: Harmonious Collaboration in Federated Learning with Competing Participants (AAAI 2024 · similarity 0.92)
- Ranking-based Client Imitation Selection for Efficient Federated Learning (ICML 2024 · similarity 0.86)
- FedCompass: Efficient Cross-Silo Federated Learning on Heterogeneous Client Devices Using a Computing Power-Aware Scheduler (ICLR 2024 · similarity 0.86)
- Federated Learning under Periodic Client Participation and Heterogeneous Data: A New Communication-Efficient Algorithm and Analysis (NeurIPS 2024 · similarity 0.86)
- Balancing Similarity and Complementarity for Federated Learning (ICML 2024 · similarity 0.85)
- Debiasing Federated Learning with Correlated Client Participation (ICLR 2025 · 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.