Hierarchical federated learning with paillier encryption: synergistic approach for secure analytics of sensitive healthcare data

Federated learning enables collaborative analysis of medical data across healthcare institutions without centralizing sensitive patient information. However, its application faces three persistent challenges, non-identically distributed data across sites, variable client resources affecting training efficiency, and privacy risks from model update exposure. We present a framework that addresses these challenges through three integrated components: hierarchical prototyping to align heterogeneous data distributions via adaptive local-global prototype weighting; dynamic client selection that evaluates participants using accuracy contribution, data diversity, and computational efficiency metrics with adaptive thresholding; and secure aggregation using the Paillier cryptosystem to encrypt model updates during transmission. Evaluation across eight healthcare benchmarks, including CURIAL, MIMIC-III, ChestMNIST, and BraTS, shows consistent improvements over established methods. On the CURIAL dataset, our method achieves 96.21% accuracy compared to 83.21% for FedAvg, representing a 13.0 percentage point gain. Concurrent improvements are observed in sensitivity (+4.1%), specificity (+3.8%), F1-score (+3.2%), AUC (+0.126), and Diagnostic Odds Ratio (+19.54). Privacy evaluation indicates 58–63% lower success rates for gradient leakage attacks and 41–47% reduced reconstruction fidelity in model inversion attempts compared to unencrypted aggregation. Ablation analysis attributes 7.2 percentage points of accuracy gain under severe non-IID conditions to hierarchical prototyping, while dynamic client selection reduces convergence oscillations by 34%. Results across imaging, tabular, and time-series modalities confirm the approach generalizes to diverse healthcare data types. Code and configuration files: https://github.com/SaeedIqbal/FedHomo