Privacy-Preserving Distributed Machine Learning And Federated Ai Systems For Healthcare And Financial Applications

Abstract

Healthcare and financial sectors increasingly rely on data-driven machine learning while facing stringent privacy, security, and regulatory constraints. Centralized learning approaches amplify data exposure risks, making them unsuitable for sensitive domains such as electronic health records and financial transactions. This study investigates privacy-preserving distributed machine learning and federated AI systems that enable collaborative model training without sharing raw data. Using experimental evaluations on healthcare-style datasets (40,000–70,000 patient records) and large-scale financial transaction data with extreme class imbalance (<2% fraud), the results show that federated learning achieves performance close to centralized models, with only 1–3% AUROC reduction. The integration of secure aggregation and differential privacy (ε ≈ 3–5) reduces membership inference and reconstruction risks by 60–85% while maintaining acceptable predictive accuracy. Robust aggregation further improves resilience under adversarial conditions, sustaining AUROC values above 0.80 even with 20% malicious participants. Overall, the findings demonstrate that privacy, robustness, and analytical utility can be jointly optimized, making federated AI a viable solution for trustworthy healthcare and financial analytics.