FedNLC: Personalized Federated Learning via Inverse-Norm Weighted Aggregation with Local Calibration

Federated learning has emerged as a crucial privacy-preserving paradigm for collaborative medical diagnosis. However, severe data heterogeneity across clinical institutions often leads to suboptimal global models. While Personalized FL via model decoupling alleviates this issue, it suffers from two critical limitations: global representation degradation caused by highly divergent local updates, and intra-client representation–classifier misalignment during the local deployment of the updated global backbone. To address these challenges, we propose FedNLC (Personalized Federated Learning via Inverse-Norm weighted aggregation with Local Calibration). Specifically, FedNLC introduces a Inverse-Norm Weighted Aggregation strategy at the server side to dynamically downweight structurally divergent updates, thereby protecting the shared global feature space from outlier contamination. Concurrently, a Mini-Batch Local Knowledge Distillation module is employed at the client side to structurally calibrate the incoming global backbone with the existing local classifier. Extensive experiments on medical imaging benchmarks demonstrate that FedNLC mitigates the decoupling dilemma, achieving superior convergence stability, local adaptability, and overall diagnostic accuracy compared to State-of-the-Art PFL methods.