Privacy-Preserving Trainer Recruitment in Model Marketplace of Federated Learning

Federated learning (FL) technologies enable trainers to collaboratively train machine learning (ML) models while maintaining data privacy, making them a crucial component of the next-generation model marketplace. However, several issues arise from the non-independent and identically distributed (non-IID) data among trainers, as well as the customers' diverse model orders. Consequently, it is necessary to design a trainer recruitment mechanism to select trainers that meet the customer's model requirements and improve the performance of purchased models. In this paper, we propose a privacy-preserving trainer recruitment scheme in a model marketplace of FL, which aims to recruit optimal trainers that meet the customer's requirements. Specifically, we propose a hierarchical recruitment mechanism to select trainers based on their task preferences, data distributions, and data sizes. Additionally, we prove the NP-hardness of the optimal trainer recruitment problem and propose a heuristic selection algorithm to provide an approximate solution. Extensive experiments demonstrate that our proposed scheme effectively improve the performance of purchased models, particularly in scenarios with highly non-IID data and limited budgets.