PrismFed: Joint Optimization via Dynamic Bayesian Persuasion for Multi-Task Federated Learning under Incomplete Information

Efficiently allocating limited communication resources in multi-task federated learning is a fundamental challenge. This challenge is particularly pronounced under incomplete information, where the server cannot observe clients' private characteristics and communication resources are scarce and dynamically varying over time. Existing approaches predominantly rely on indirect information completion or inference techniques, which are often unstable and insufficiently adaptive in time-varying environments, thereby limiting performance gains. To address these challenges and effectively incentivize client participation, we propose PrismFed, a joint optimization framework via dynamic Bayesian persuasion for multi-task federated learning under incomplete information. The interaction between the server and the clients is formulated as a dynamic Bayesian persuasion game, which enables the server to strategically influence participation decisions while coordinating communication resource allocation across tasks. We characterize the optimal dynamic persuasion policy and theoretically establish its effectiveness in guiding participation under uncertainty. Building on this policy, we further derive the per-round optimal communication resource allocation across multiple tasks, jointly optimizing system utility in dynamic environments. Extensive experiments show that PrismFed improves task-weighted accuracy and system revenue across two task combinations, achieving gains of 2.51% over the strongest baseline.