PIECE: Incentivizing Personalized Privacy-Preserving for Multi-Version Model Marketplace in Federated Learning

Although Federated Learning (FL) offers significant potential for developing model marketplaces through collaborative training and privacy preservation, challenges such as insufficient training data and arbitrage issues severely impede the development of FL-based model marketplaces. Existing studies either lack satisfactory security guarantees or are too profit-driven to address potential arbitrage issues. In this paper, we propose a novel Personalized prIvacy-prEserving inCentive mEchanism named PIECE, with the aim of achieving social optimality while avoiding arbitrage. Specifically, we first formulate a dual-objective optimization problem to simultaneously maximize social utility and model performance while ensuring arbitrage-free conditions through differential privacy. Due to dynamic model training and heterogeneous privacy budgets that complicate the design of arbitrage-free properties, we model the transformation between local and global privacy requirements across scenarios as a privacy choice game. This game guarantees the identification of a constraint to generate desired model versions based on Nash equilibrium. Next, by generalizing the properties of different data-owner groups under equilibrium conditions, we prove that the dual-objective optimization problem is always conflict-free, thus allowing transformation into a social optimal problem without arbitrage. Furthermore, to tackle the significant difficulty in characterizing the model revenue and interpolating pricing, we propose a two-stage solution based on subadditivity relaxation. The first stage establishes a set of ideal prices as the target, while the second stage establishes polynomial-time solvability and provides rigorous arbitrage-free boundaries. Finally, comprehensive experiments on four real-world datasets validate the efficacy of PIECE. The results indicate a minimum 8% boost in model revenue within the specified marketplace scale, and a maximum 16.67% improvement in model performance compared to the state-of-the-art baselines.