Balancing Privacy And Utility: Strategies For Differential Privacy In Healthcare Machine Learning Models

Proceedings of the 65th International Association for Computer Information Systems Conference - October 1 - 4, 2025 | Clearwater 54 BALANCING PRIVACY AND UTILITY: STRATEGIES FOR DIFFERENTIAL PRIVACY IN HEALTHCARE MACHINE LEARNING MODELS Nick Pierrelouis, Marymount University, [email protected]du Xiang Liu, Marymount University, [email protected] PROPOSED STUDY This study addresses the critical challenge organizations face when implementing differential privacy (DP) in AI-enabled Cardiac Monitoring Wearable Devices (AICMWD). As healthcare institutions increasingly deploy Machine Learning Healthcare Models (MLHM) to process sensitive patient data, they must balance strong privacy protections with model utility (Dwork & Roth, 2013). This research is relevant to IACIS participants as it provides actionable implementation strategies for organizations seeking to leverage Artificial Intelligence (AI) innovations, while also adhering to regulatory requirements and fostering patient trust in an era where health data breaches have become increasingly common and costly (Vallepu, 2024). BASIS OF THE STUDY The research employs the Investigation, Implementation, and Assessment (IIA) methodology to evaluate DP implementation across the ML lifecycle. Data collection focuses on synthetic datasets simulating Food and Drug Administration (FDA)-regulated cardiac monitoring devices, with analysis conducted during the model deployment and inference monitoring under the CRossIndustry Standard Process model for the development of Machine Learning applications with Quality assurance methodology (CRISP-ML(Q)) framework (Studer et al., 2021). The study systematically adjusts privacy parameters ϵ (epsilon values) to determine optimal configurations that protect against membership and attribute inference attacks in a black box setting (Wu et al., 2024). The research evaluates how DP implementation decisions impact security, operational performance, and privacy-utility trade-offs in healthcare AI systems. Preliminary findings reveal that strategic implementation of DP mechanisms at specific ML pipeline stages can significantly reduce utility loss while ensuring strong privacy protections (Abadi et al., 2016). This positions organizations for operational success by enabling practical deployment guidelines based on their risk tolerance and performance requirements. IMPLICATIONS The findings may significantly impact organizational privacy governance in MLHD management and security. With increasing regulatory mandates on privacy-preserving AI/ML medical devices, organizations must develop structured DP approaches without compromising clinical effectiveness (Biasin et al., 2023). This research demonstrates that organizations can achieve compliance without sacrificing innovation by adopting the IIA methodology and selecting appropriate privacy parameters based on data sensitivity and attack vectors. Additionally, results highlight the need for governance frameworks explicitly addressing ML lifecycle deployment and maintenance phases, where inference attacks pose the highest privacy risk (Vizitiu et al., 2021). Proceedings of the 65th International Association for Computer Information Systems Conference - October 1 - 4, 2025 | Clearwater 55 CONCLUSIONS This research concludes that organizations can effectively balance privacy and utility in healthcare machine learning by strategically implementing DP mechanisms. Instead of applying uniform privacy approaches, organizations should conduct systematic risk assessments of potential inference attacks and deploy targeted protections at critical ML pipeline stages (Sahiner et al., 2023). The IIA methodology serves as a structured framework for evaluating, implementing, and continuously assessing privacy-preserving measures in healthcare AI applications, ensuring regulatory compliance and clinical effectiveness. Integrating privacy into AI governance enables healthcare organizations to leverage ML innovations responsibly while preserving patient trust and data security (Cummings et al., 2024). REFERENCES Abadi, M., Chu, A., Goodfellow, I., McMahan, H. B., Mironov, I., Talwar, K., & Zhang, L. (2016). Deep Learning with Differential Privacy. Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security, 308–318. https://doi.org/10.1145/2976749.2978318 Biasin, E., Kamenjasevic, E., & Ludvigsen, K. R. (2023). Cybersecurity of AI medical devices: Risks, legislation, and challenges (No. arXiv:2303.03140). arXiv. https://doi.org/10.48550/arXiv.2303.03140 Cummings, R., Desfontaines, D., Evans, D., Geambasu, R., Huang, Y., Jagielski, M., Kairouz, P., Kamath, G., Oh, S., Ohrimenko, O., Papernot, N., Rogers, R., Shen, M., Song, S., Su, W., Terzis, A., Thakurta, A., Vassilvitskii, S., Wang, Y.-X., … Zhang, W. (2024). Advancing Differential Privacy: Where We Are Now and Future Directions for Real-World Deployment. Harvard Data Science Review, 6(1). https://doi.org/10.1162/99608f92.d3197524 Dwork, C., & Roth, A. (2013). The Algorithmic Foundations of Differential Privacy. Foundations and Trends® in Theoretical Computer Science, 9(3–4), 211–407. Sahiner, B., Chen, W., Samala, R. K., & Petrick, N. (2023). Data drift in medical machine learning: Implications and potential remedies. The British Journal of Radiology, 96(1150), 20220878. https://doi.org/10.1259/bjr.20220878 Studer, S., Bui, T. B., Drescher, C., Hanuschkin, A., Winkler, L., Peters, S., & Mueller, K.-R. (2021). Towards CRISP-ML(Q): A Machine Learning Process Model with Quality Assurance Methodology (No. arXiv:2003.05155). arXiv. http://arxiv.org/abs/2003.05155 Vallepu, R. (2024). Exploring Data Security and Privacy Challenges in Master Data Governance Systems. International Journal of Computer Trends and Technology, 72(11), 126–134. https://doi.org/10.14445/22312803/IJCTT-V72I11P113 Vizitiu, A., Nita, C.-I., Toev, R. M., Suditu, T., Suciu, C., & Itu, L. M. (2021). Framework for Privacy-Preserving Wearable Health Data Analysis: Proof-of-Concept Study for Atrial Fibrillation Detection. Applied Sciences, 11(19), Article 19. https://doi.org/10.3390/app11199049 Wu, F., Cui, L., Yao, S., & Yu, S. (2024). Inference Attacks: A Taxonomy, Survey, and Promising Directions (No. arXiv:2406.02027). arXiv. https://doi.org/10.48550/arXiv.2406.02027