Robust Federated Learning for Heterogeneous Model and Data.

Area of Science:

• Medical Informatics
• Machine Learning
• Data Security

Background:

• Data privacy and security are critical challenges in clinical settings with sensitive patient data.
• Federated Learning (FL) enables collaborative model training across hospitals without sharing raw data.
• Heterogeneity in data and models across institutions poses significant challenges to robust FL, including gradient leakage risks.

Purpose of the Study:

• To propose a robust Federated Learning (FL) method for tackling data and model heterogeneity in clinical settings.
• To enhance the security and accuracy of collaborative machine learning models using sensitive medical data.
• To lay the foundation for reliable FL applications in laboratories and clinical practice.

Main Methods:

• Developed a novel FL method incorporating knowledge distillation and a weighted client confidence score.
• Trained models on hematological cytomorphology data to address clinical data challenges.
• Utilized symmetric loss to mitigate the impact of data heterogeneity and noisy labels.

Main Results:

• The proposed FL method demonstrated superior performance compared to existing approaches.
• Successfully addressed both data and model heterogeneity in an end-to-end FL framework.
• Showcased effective knowledge transfer from clean models to others, mitigating issues from noisy clients.

Conclusions:

• The developed FL method offers a robust solution for heterogeneous medical data and models.
• This work is the first to address data and model heterogeneity comprehensively in end-to-end FL for clinical applications.
• The findings pave the way for more secure and effective FL deployment in healthcare and research.