Budget-Feasible Double Auction Mechanisms for Model Training Services in Federated Learning Market

In cross-silo Federated Learning (FL), where data is distributed across multiple organizations such as hospitals, incentivizing collaboration while preserving data privacy presents a significant challenge. FL markets facilitate transactions among these organizations, introducing a budget-feasible two-sided allocation problem. Budget constraints are a realistic and crucial issue that reflects the economic capacities and engagement levels of market participants, but they also increase the complexity of allocation. To address this challenge, this paper proposes BudoMech, a novel budget-feasible double auction mechanism. BudoMech integrates requesters’ values and providers’ costs to determine allocations and payments, ensuring adherence to budget constraints while guaranteeing truthfulness. The mechanism sorts values and costs, uses virtual prices to iteratively calculate allocation quantities and prices, and then outputs the final allocations and payments. Theoretical analysis confirms that BudoMech satisfies desirable properties, including budget feasibility, individual rationality, truthfulness, and efficiency. Furthermore, extensive experiments demonstrate that BudoMech is an effective solution for resource allocation in FL markets.

I. Introduction

In recent years, Federated Learning (FL) [1], [2] has emerged as a powerful paradigm for enabling collaborative model training across distributed data sources without the need to share private raw data. One specific type of FL, known as cross-silo FL [3], [4], involves collaboration among distinct organizations, such as hospitals, financial institutions, or government agencies. In this context, each organization participant operates as a client in the FL process, contributing data and computational resources to the collective model while maintaining data privacy. For a signal organization, the data heterogeneity and limited data availability can lead to unsatisfied model performance when training is conducted solely on its own data. To enhance model performance, organizations may seek to engage in collaborative efforts through FL. However, organizations in cross-silo FL, such as hospitals, often hold valuable private data and computing resources that they are reluctant to share due to privacy concerns or competitive reasons. Additionally, some organizations may hesitate to participate without financial or other incentives [5]. This highlights the need for mechanisms that not only address privacy and security concerns but also provide incentives to encourage active participation in FL.