A Federated Learning Framework with Differential Privacy for High-Accuracy and Privacy-Preserving Healthcare Prediction

Abstract

The proliferation of sensitive patient data across distributed healthcare institutions has created an urgent need for predictive models that safeguard privacy without sacrificing diagnostic accuracy. We propose a federated learning framework integrated with differential privacy to address this challenge in intelligent healthcare prediction. Our approach enables multiple hospitals and clinics to collaboratively train a deep neural network—such as a Multi-Layer Perceptron or a Transformer-based encoder—without ever sharing raw clinical or demographic records. Each participating institution performs local data preprocessing and model training on its own premises, thereby ensuring that no identifiable information leaves the local infrastructure. To further strengthen privacy guarantees, we inject calibrated Gaussian noise into each client’s model updates before transmission to the central server. This perturbation obscures the contribution of any single patient record, making it computationally infeasible for an adversary to reconstruct sensitive information from the communicated gradients. The server then aggregates these privatized updates using a weighted Federated Averaging algorithm, where each institution’s influence on the global model is proportional to its local dataset size. The resulting global model is redistributed to all clients for subsequent training rounds, and this iterative process continues until convergence. Our framework is specifically designed to handle non-independent and identically distributed (non-IID) data across institutions, a common challenge in real-world healthcare settings. The primary novelty lies in the seamless combination of local training, differential privacy, and weighted aggregation to achieve high predictive accuracy while maintaining rigorous privacy standards. This work demonstrates that strong privacy protection and state-of-the-art prediction performance are not mutually exclusive, thereby offering a practical and scalable solution for collaborative healthcare analytics.