FULL PDF
Abstract
Federated Learning (FL) enables collaborative model training across distributed data silos while preserving data privacy—a particularly appealing approach for healthcare, where data sensitivity and institutional silos dominate. However, the deployment of FL in medical domains is challenged by data heterogeneity and varying data quality across institutions. This paper presents a comprehensive review and introduces optimizations to FL protocols to address these challenges. Specifically, we investigate adaptive client weighting, quality-aware aggregation, and robust differential privacy schemes. Our analysis shows that these strategies not only maintain model performance but also ensure fairness and privacy across clients with diverse data characteristics.
View more >>