Weighted Federated Learning with encryption for diabetes classification

This study presents an innovative weighted Federated Learning (FL) framework with integrated encryption for diabetes classification across multiple healthcare institutions. Our comprehensive approach addresses three critical challenges in collaborative healthcare analytics: data privacy preservation, non-IID data distribution, and model performance optimization. The framework incorporates a weighted aggregation mechanism based on local data volumes to effectively handle client data imbalance, while implementing a lightweight masking-based encryption scheme to protect model parameters during transmission without compromising computational efficiency. We evaluate our approach using a comprehensive dataset of 15,347 entries from three internationally recognized medical organizations (ADCES, CDC, IDF) across five machine learning models: Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), three-layer Deep Neural Network (DNN), and deeper five-layer network (Deeper DNN). Experimental results demonstrate that weighted FL consistently equals or surpasses centralized learning performance while maintaining strict privacy compliance. Notable improvements include SVM AUC enhancement from 0.46 to 0.57 and RF AUC improvement from 0.70 to 0.76. The encryption mechanism introduces negligible overhead (0.0001s encryption, 0.0013s decryption per round) with minimal communication costs (0.16 KB per round). Our framework’s ability to securely handle non-IID healthcare datasets while providing interpretable results through SHAP analysis positions it as a practical solution for privacy-preserving collaborative diagnostics. This research represents a significant advancement toward scalable, privacy-conscious medical analytics that can be adopted across diverse healthcare institutions without compromising data sovereignty or diagnostic accuracy.