Personalized-Enhanced Federated Learning on Heterogeneous Internet of Medical Things

The significant heterogeneity of data resources in Internet of Medical Things (IoM
$T$
) devices profoundly affects the efficacy of federated learning (FL) when training medical models. Personalized Federated Learning (pFL) not only facilitates the extraction of shared priors from a vast array of distributed data in medical diagnostics but also underpins the development of bespoke models for heterogeneous medical devices. However, existing personalized federated learning methods have problems with insufficient personalization and poor convergence in medical data.To address this issue, we propose an Enhanced Personalized Federated Learning (EPFL) method, designed specifically for heterogeneous IoMT federated learning. Considering the diverse knowledge characteristics of model parameters, we introduce the concept of frequency domain analysis to deeply scrutinize the composition of model parameters. This approach enables us to accurately distinguish between shared and personalized features within the medical model. Moreover, we leverage the principles of knowledge transfer and cooperative gaming to achieve person-alized effects in medical models. We have conducted extensive experiments using various datasets in medical scenarios charac-terized by heterogeneous data. The results clearly demonstrate that, in comparison to current state-of-the-art benchmarks, EPFL achieves superior personalized accuracy for each medical device.