Privacy-Preserving of Hospital Data using Secure Multi-Party Computation and Federated Neural Models with Encrypted Inference Aggregation
Yuvraj Pandey, Devansh Agarwal, Sheryl Oliver, M. Kiruthika · 2026
In this study, we propose a machine learning framework based on SMPC for privacy-preserving secure model training and inference in a federated manner. Residual learning was used to discover potential direct mappings, enabling the application of a feed-forward neural network (FFNN) over M = 10 data owners while the privacy of data was retained via encryption schemes, differential privacy, and secure aggregation. No sensitive information was exposed during training or inference, and the achieved accuracy for predicting outcome was 80.6%. Differential privacy with Laplace noise was used achieving strong privacy maintenance with slight degradation in performance. We show through an experimental validation that we can perform aggregated secure encrypted predictions without loss of accuracy. Federated learning combined with SMPC offers a scalable, privacy-preserving solution for distributed AI applications. In future research, we will aim not only at improving efficiency in terms of computation but also strengthen the privacy-preserving strategies, while exploring more efficient model-based inference techniques, such as homomorphic encryption, in which the model could be encrypted entirely.