Knowledge-Free Correlated Agreement for Incentivizing Federated Learning

A team of researchers led by Leon Witt (including Togrul Abbasli, Kentaroh Toyoda, Wojciech Samek, and Lucy Klinger) has published a paper on arXiv introducing Knowledge-Free Correlated Agreement (KFCA), a novel incentive mechanism for federated learning (FL). The core challenge in FL is fairly rewarding clients who contribute training data without having access to a central test set or ground-truth labels. Existing methods like Correlated Agreement (CA) are vulnerable to label-flipping attacks where malicious clients manipulate their reports. KFCA solves this by being strictly truthful under the assumption of an honest majority and requiring only categorical reports from clients—no public test set, no distribution knowledge, and no ground-truth labels needed. This makes it ideal for privacy-sensitive and decentralized FL scenarios.

KFCA was evaluated on two realistic tasks: federated fine-tuning of large language model (LLM) adapters and a real-world printed circuit board (PCB) inspection dataset. The results demonstrated efficient real-time reward computation, a critical feature for blockchain-based or decentralized FL systems where compensation must be immediate and low-overhead. The paper highlights KFCA's applicability in environments where trust is minimal and incentive alignment is crucial—e.g., cross-silo FL among competing enterprises or decentralized sensor networks. By removing the need for a trusted third party to hold test sets, KFCA opens the door to truly permissionless FL marketplaces.