Research map / Federated learning and privacy / Privacy attacks and secure learning
FedCompetitors: Harmonious Collaboration in Federated Learning with Competing Participants
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 FedCompetitors, a graph-theoretic and ILP-based polynomial-time collaborator selection framework for federated learning participants that compete, enforcing "the friend of my enemy is my enemy" to avoid conflicts of interest.
Problem
Selecting collaborators for federated learning participants when data is heterogeneous and some participants are competitors, so that no participant indirectly helps its competitor.
Approach
Extends the balance-theory principle "the friend of my enemy is my enemy" into an FL constraint on a data usage graph (no competitor may be reachable via collaborations), formulated with graph theory and ILP and solved by a greedy O(n^3) per-participant algorithm using reachability matrices.
Key results
- On CIFAR-10 with random competing graphs, FedCompetitors reaches 91.33 MTA versus 87.80 for CE (about 3.53 points higher), with FedAvg, FedProx and SCAFFOLD around 51-53.
- In the weakly Non-IID synthetic setting, FedCompetitors achieves the lowest MSE on nearly all FL-PTs (e.g., 0.13-0.15), while CE is worse on small FL-PTs (about 1.1-1.2 MSE).
- On eICU, FedCompetitors attains the best AUC for most of the 10 hospitals (e.g., 95.24 for v10 and 89.58 for v6), though it is not best for v1 and v3.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedCompetitors instead of CE | CIFAR-10 FL with randomly generated competing graphs · CIFAR-10 | MTA: +3.53% | Helped |
| Use FedCompetitors instead of the baselines | CIFAR-10 FL with randomly generated competing graphs · CIFAR-10 | MTA: 91.33 ± 4.14; CE: 87.80 ± 7.18 | Helped |
| Use FedCompetitors instead of CE | eICU hospital FL · eICU | AUC | Mixed |
| Use FedCompetitors instead of CE | Synthetic weakly Non-IID FL · Synthetic weakly Non-IID | MSE | Mixed |
| Use FedCompetitors instead of FedAvg, FedProx, or SCAFFOLD | Synthetic strongly Non-IID FL · Synthetic strongly Non-IID | MSE | Helped |
| Use FedCompetitors instead of the five baselines | Synthetic strongly Non-IID FL · Synthetic strongly Non-IID | MSE | Helped |
| Use FedCompetitors instead of the baselines | eICU hospital FL · eICU | AUC | Helped |
Surprising findings
- Local training outperforms FedAvg, FedProx and SCAFFOLD in the strongly Non-IID synthetic setting, since a single global model cannot satisfy conflicting tasks.
- In the weakly Non-IID synthetic setting, CE performs worst on average because small FL-PTs cannot benefit from large ones.
Limitations
- Experiments use small numbers of FL-PTs (8 to 10), limiting evidence of scalability.
- Benefit graph assumed to be given or obtained via a hypernetwork technique from prior work; quality depends on it.
- Greedy algorithm solves the ILP heuristically, with no optimality guarantee stated in the excerpt.
Experimental setup
Datasets: synthetic polynomial regression data CIFAR-10 eICU
Benchmarks: synthetic polynomial regression CIFAR-10 eICU
Metrics: MSE mean test accuracy (MTA) AUC
Baselines: Local FedAvg FedProx SCAFFOLD CE
Builds on: FedAvg FedProx SCAFFOLD collaboration equilibrium balance theory
Closest papers by meaning
- Free-Rider and Conflict Aware Collaboration Formation for Cross-Silo Federated Learning (NeurIPS 2024 · similarity 0.92)
- Balancing Similarity and Complementarity for Federated Learning (ICML 2024 · similarity 0.89)
- Bridging Model Heterogeneity in Federated Learning via Uncertainty-based Asymmetrical Reciprocity Learning (ICML 2024 · similarity 0.88)
- Ranking-based Client Imitation Selection for Efficient Federated Learning (ICML 2024 · similarity 0.88)
- A Bayesian Approach for Personalized Federated Learning in Heterogeneous Settings (NeurIPS 2024 · similarity 0.87)
- Scalable and Trustworthy Learning in Heterogeneous Networks (AAAI 2025 · similarity 0.87)
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.