PrivFL-SMPC: Privacy-aware federated learning using secure multi-party computation for collaborative diabetes diagnosis in the cloud

Collaborative analytics across multiple healthcare institutions has become increasingly important for developing accurate predictive models while preserving patient privacy. Conventional centralized machine learning approaches require data sharing, creating significant privacy, security, and regulatory challenges. To address these issues, this paper proposes PrivFL-SMPC, a privacy-aware federated learning framework that enables collaborative healthcare analytics on structured clinical datasets without exposing sensitive patient information. The framework performs data preprocessing, including cleaning, normalization, and categorical encoding, followed by local model training using the LightGBM classifier optimized through a Hybrid Red Kite–Runge Kutta Optimizer (HKRO) to improve predictive performance and computational efficiency. To protect model updates during collaborative training, the framework integrates Additive Secret Sharing (ASS) for secure distribution of local model parameters and the Paillier cryptosystem for privacy-preserving secure aggregation. This design enables participating healthcare institutions to collaboratively train a global model without directly sharing raw patient data or intermediate model updates. Experimental evaluation on a diabetes dataset demonstrates that PrivFL-SMPC achieves an accuracy of 0.9726, precision of 0.97, recall of 0.96 and an F1-score of 0.96, while maintaining the confidentiality of model updates throughout the collaborative learning process. The proposed framework provides an effective and scalable solution for privacy-preserving healthcare analytics in distributed cloud environments.