Mechanism for Decision-aware Collaborative Federated Learning: A Pitfall of Shapley Values

This paper investigates mechanism design for decision-aware collaboration via federated learning (FL) platforms. Our framework consists of a digital platform and multiple decision-aware agents, each endowed with proprietary data sets. The platform offers an infrastructure that enables access to the data, creates incentives for collaborative learning aimed at operational decision-making, and conducts FL to avoid direct raw data sharing. The computation and communication efficiency of the FL process is inherently influenced by the agent participation equilibrium induced by the mechanism. Therefore, assessing the system's efficiency involves two critical factors: the surplus created by coalition formation and the communication costs incurred across the coalition during FL. To evaluate the system efficiency under the intricate interplay between mechanism design, agent participation, operational decision-making, and the performance of FL algorithms, we introduce a multi-action collaborative federated learning (MCFL) framework for decision-aware agents. Under this framework, we further analyze the equilibrium for the renowned Shapley value based mechanisms. Specifically, we examine the issue of false-name manipulation, a form of dishonest behavior where participating agents create duplicate fake identities to split their original data among these identities. By solving the agent participation equilibrium, we demonstrate that while Shapley value effectively maximizes coalition-generated surplus by encouraging full participation, it inadvertently promotes false-name manipulation. This further significantly increases the communication costs when the platform conducts FL. Thus, we highlight a significant pitfall of Shapley value based mechanisms, which implicitly incentivizes data splitting and identity duplication, ultimately impairing the overall efficiency in FL systems.