FCH, an incentive framework for data-owner dominated federated learning

Abstract

Federated learning has gained prominence for its superior privacy-preserving properties. However, establishing an incentive framework that motivates participants to contribute fully is essential to mitigate opportunistic behaviour arising from information asymmetry. Current frameworks primarily incentivize data owners to contribute resources in scenarios where model owners are dominant. This article addresses the underexplored scenario where data owners are in charge and the potential unethical behaviours it may entail. We propose a new framework, the Fair Clearing House (FCH), that promotes balanced participation from data owners and model owners, optimizing their contributions to the learning process. Numerical results demonstrate that FCH outperforms existing frameworks under comparable conditions while reducing unethical behaviours.