HPFL: A Hierarchically Personalized Federated Learning Framework with FedMARL-Based Mediator Selection

Federated Learning (FL) enables collaborative model training across distributed clients without sharing raw data, but conventional methods like FedAvg often fail under non-IID data settings due to limited personalization and equal-weight aggregation. We propose a novel Hierarchically Personalized Federated Learning (HPFL) framework that introduces an intermediate layer of me- diator servers—each linked to a specific client node. These mediators inherit the central model’s weights but are updated preferentially by their corresponding node’s gradients, enabling enhanced local personalization. To further improve global model performance and fairness, we integrate FedMARL-based me- diator selection[9], where a Multi-Agent Reinforcement Learning (MARL) framework is employed to evaluate and select the most informative mediator models at each round. Only these top- performing mediators are used to update the global model, ensuring adaptive aggregation that favors high-quality, well-personalized contributions. This structure mirrors the agent-mediator- controller hierarchy found in MARL systems and enhances scalability without requiring direct MARL integration at the client level. Our approach aligns with existing research in robust privacy- preserving techniques to prevent poisonous attacks [7], model aggregation methods[8], supported by previous surveys related medical image analysis[10].