FedMGKD: a multi-granularity trusted knowledge distillation framework for edge personalized federated learning

In edge computing environments, Personalized Federated Learning (PFL) has garnered significant attention due to its strong capabilities in preserving data privacy and enabling local adaptation. However, existing knowledge distillation-based PFL approaches often suffer from limited transfer efficiency and robustness, particularly when confronted with heterogeneous data distributions and erroneous knowledge propagation. To address these challenges, this paper proposes FedMGKD, a novel federated learning framework based on multi-granular trustworthy knowledge distillation, aiming to achieve an optimal balance between global generalization and local customization. Specifically, FedMGKD aligns intermediate representations between student and teacher models across multiple layers of a global feature extractor—including embedding and intermediate convolutional layers—to capture richer shared semantics. Moreover, an exponential moving average correction mechanism is introduced to adaptively regulate the distillation strength based on teacher confidence, thereby mitigating the impact of unreliable knowledge. To further disentangle global and local features, an orthogonal constraint is applied, ensuring the preservation of core shared information while maintaining client-specific representations. Extensive experiments demonstrate that FedMGKD consistently outperforms state-of-the-art methods across various non-IID data distributions and client-scale scenarios. It achieves significant improvements in accuracy and robustness, all while maintaining low communication overhead and strong privacy guarantees. The code is available at https://github.com/lwl-525/FedMGKD .