At What Price? Exploring the Potential and Challenges of Differentially Private Machine Learning for Healthcare - Proceedings of the 56th Hawaii International Conference on System Sciences 2023

Aug 3, 2026

Upload by Jiven

The growing availability of healthcare data accelerates innovation in machine learning, but increased processing of sensitive information intensifies privacy risks. Differential privacy offers a mechanism by introducing randomness to protect individuals, yet it typically degrades model accuracy, creating a privacy–utility trade-off. This work provides an empirical evaluation of two differentially private learning approaches on medical image and text data, clarifying potential benefits and practical limitations for healthcare deployments.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages10

Reading Time25 min

AccessFree to download

Abstract

Introduction

Machine learning drivers in healthcare

Data privacy issues in ML

Differential privacy as a solution and its trade-off

Proceedings of the 56th Hawaii International Conference on System Sciences | 2023 At What Price? Exploring the Potential and Challenges of Differentially Private Machine Learning for Healthcare Aycan Aslan University of Goettingen [aycan.aslan@uni-goettingen.de](aycan.aslan@uni-goettingen.de) Tizian Matschak University of Goettingen [tizian.matschak@uni-goettingen.de](tizian.matschak@uni-goettingen.de) Maike Greve University of Goettingen [maike.greve@uni-goettingen.de](maike.greve@uni-goettingen.de) Simon Trang University of Goettingen [strang@uni-goettingen.de](strang@uni-goettingen.de) Lutz M. Kolbe University of Goettingen [lkolbe@uni-goettingen.de](lkolbe@uni-goettingen.de) Abstract The increased generation of data has become oneof the main drivers of technological innovation in healthcare. This applies in particular to the adoption of Machine Learning models that are used to generate value from the growing available healthcare data. However, the increased processing of sensitive healthcare data comes with challenges in terms of data privacy. Differential privacy, the method of adding randomness to the data to increase privacy, has gained popularity in the last few years as a possible solution. However, while the addition of randomness increases privacy, it also reduces overall model performance, generating a privacy-utility trade-off. Examining this trade-off, we contribute to the literature by providing an empirical paper that experimentally evaluates two prominent and innovative methods of differentially private Machine Learning on medical image and textdata to deepen the understanding of the existing potential and challenges of such methods for the healthcare domain. Keywords: Differential privacy, PATE framework, Differentially private stochastic gradient descent. 1. Introduction The digitization of healthcare data and technological advancements in computer processing and data storage has enabled the development of advanced algorithmic techniques such as Artificial Intelligence, especially in the form of Machine Learning (ML) . Beyond the increase in volume, velocity, and variety of available healthcare data, an additional driver of ML applications is financial pressures on the healthcare industry globally, with increasing demands due to a growing and aging population (Stanfill & Marc, 2019) . Against this background, the use of ML is gaining popularity not only in research but also in medical practices. For instance, to realize the potential offered by ‘precision medicine’, a tailored medical treatment of patients based on individual characteristics (Ginsburg & Phillips, 2018), data from a wide range of data sources, such as Electronic Health Records (EHR) or genomics data, must be collected and subsequently analyzed (Ginsburg & Phillips, 2018) . Here, the high speeds at which ML models perform make them a suitable tool to efficiently take advantage of a growing, diverse set of healthcare data (Jiang et al., 2017) . For example, studies show that ML models can be used to analyze gene expression data and DNA data to predict the treatment response of patients with rheumatoid arthritis (Tao et al., 2021) . Other examples of the use of ML show its potential for an automated system of disease classification of medical images (Mehta & Pandit, 2018) and fraud detection within the healthcare system (Matschak et al., 2022) . While the presented examples illustrate the potential of ML for healthcare, one cannot ignore the associated data privacy issues. These issues primarily originate from the high demand ML places on vast amount of data to train on, but there are also privacy issues arising from the inherent nature of ML (Abouelmehdi et al., 2018) . These include the privacy of model weights of ML models or the possible data memorization of individual data points during the training of ML models (Kaissis et al., 2020) . Studies have shown that, for example, model inversion attacks can be performed

FAQs

Why is differential privacy used in healthcare machine learning?

Differential privacy reduces the risk of retrieving information about individuals from sensitive healthcare data, making privacy protection more quantifiable and applicable to ML systems.

What problem does differential privacy introduce for machine learning models?

Adding randomness to achieve privacy protection reduces overall model performance, creating a privacy–utility trade-off.

What does the paper evaluate experimentally?

It experimentally compares two prominent differentially private machine learning methods on medical image and text data to study both potential and challenges in the healthcare domain.

Aug 3, 2026

Upload by Jiven