Domination Strategies for Free-Riding in Cross-Silo FL-based Caching

Federated Learning (FL) has been widely applied to content popularity prediction in caching systems to enhance Quality of Experience (QoE). In cross-silo FL, distributed organizations with local caches collaboratively train a global model coordinated by a central server. As the global model is a public good, caches may benefit without contributing, leading to free-riding. Since local request data is time-varying, repeated FL tasks form a long-term game in which free-riding destabilizes social welfare, defined as the sum of expected payoffs of all caches. This paper proposes a multi-player multi-action zero-determinant (MMZD) strategy to dominate free-riders and control social welfare. Using the FedML framework and YouTube datasets, experiments show that MMZD achieves stable and high social welfare.