CAST: Cluster-Driven Truthful Crowdfunding Mechanism for Shared AI Service Deployment

The rapid growth of AI models has substantially increased the cost of deploying services on servers. Fortunately, their generality enables them to serve multiple clients with similar tasks, presenting opportunities to optimize service deployment and maximize social utility by service sharing. Existing approaches for shared service deployment often rely on centralized algorithms that overlook privacy concerns and selfish behavior, leading to untruthful and suboptimal outcomes in distributed networks. While Vickrey-Clarke-Groves (VCG) mechanisms are widely used to ensure truthfulness, their adaptation to shared service deployment faces challenges related to individual rationality and scalability. To tackle these issues, we propose a novel auction mechanism based on the VCG framework, enhanced with targeted clustering algorithms. Theoretical analysis demonstrates that the clustering algorithms with controllable radius bound can mitigate the limitations of naive VCG. To preserve privacy, we further design a Bayesian, utility-likelihood clustering scheme that requires clients to reveal only utility values rather than individual details. By combining VCG with adaptive clustering, our mechanism achieves guarantees in truthfulness, economic properties, and performance. Experimental evaluations, including a case study in Non-IID federated learning, validate the effectiveness of the proposed approach, showing significant improvements in overall utility and reduction in negative individual utility instances by over 90%.