Knowledge-Free Correlated Agreement for Incentivizing Federated Learning
Researchers have introduced a novel method called Knowledge-Free Correlated Agreement (KFCA) aimed at encouraging client participation in federated learning (FL) without needing any ground truth, a public test set, or specific knowledge about data distribution. KFCA ensures that truthfulness is maintained when handling categorical inputs, especially with a majority of honest contributions, which reduces the risk of label flipping found in the existing Correlated Agreement (CA) technique. This method was tested in scenarios like federated LLM adapter tuning and practical PCB inspections, proving effective for real-time reward calculations, making it well-suited for decentralized and blockchain-based incentive systems. You can find the research paper on arXiv.
Key facts
- KFCA rewards client contributions in federated learning without ground truth, public test set, or distribution knowledge.
- KFCA is strictly truthful under categorical reports and an honest majority.
- KFCA addresses the label-flipping vulnerability of Correlated Agreement (CA).
- KFCA was evaluated on federated LLM adapter tuning and a real-world PCB inspection task.
- KFCA enables efficient real-time reward computation.
- KFCA is suitable for decentralized and blockchain-based incentive designs.
- The paper is published on arXiv.
- The paper is categorized under Computer Science > Machine Learning.
Entities
Institutions
- arXiv