A comparative synthesis of federated learning and differential privacy synergy: cross-domain insights from finance and healthcare deployments

Abstract

The simultaneous demand for predictive accuracy and strict confidentiality has positioned federated learning (FL) and differential privacy (DP) as the two leading paradigms for privacy-preserving artificial intelligence. While each has been studied extensively in isolation, the operational synergy between them — and the way that synergy plays out under finance and healthcare regulatory regimes — remains under-characterized. This paper develops a comparative study of FL and DP cooperation across the two domains, focusing on three points of friction: how the privacy budget ε interacts with model utility under cross-silo communication, how communication overhead shapes the choice of DP composition strategy, and how compliance obligations from HIPAA, GLBA, and the GDPR shift the deployment surface. We construct a controlled experimental scaffold using a synthetic financial credit-default corpus and a synthetic hospital electronic-health-record corpus, design five competing pipelines, and evaluate them on a unified set of utility, privacy, and communication metrics. FL combined with central-DP achieved the best utility–privacy balance for healthcare risk prediction, while FL with local-DP showed a wider utility gap on the financial side under class imbalance. The study positions FL–DP synergy as an engineering design space rather than a single fixed recipe.