Privacy, Utility, Effort, Transparency and Fairness - Identifying and Swaying Trade-offs in Privacy Preserving Machine Learning through Hybrid Methods

Aug 4, 2026

Upload by Eden

As Artificial Intelligence expands across economic sectors, Privacy Preserving Machine Learning (PPML) becomes essential for handling sensitive data throughout machine learning pipelines. Each PPML method protects privacy but introduces measurable trade-offs that practitioners need to understand. The work uses design science research supported by an extensive literature review to collect trade-off dimensions and method impacts. It then evaluates trade-offs via an expert focus group and synthesizes an overview of PPML methods and hybrid combinations’ effects across privacy, utility, effort, transparency, and fairness.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages15

Reading Time38 min

AccessFree to download

Abstract

Introduction

Theoretical foundation

PPML method categories

Data publishing approaches

Data processing approaches

Privacy, Utility, Effort, Transparency and Fairness: Identifying and Swaying Trade-offs in Privacy Preserving Machine Learning through Hybrid Methods Marian Eleks1, Jakob Ihler2, Jonas Rebstadt3, Henrik Kortum-Landwehr4 and Oliver Thomas5 Abstract: As Artificial Intelligence (AI) permeates most economic sectors, the discipline Privacy Preserving Machine Learning (PPML) gains increasing importance as a way to ensure appropriate handling of sensitive data in the machine learning process. Although PPML-methods stand to provide privacy protection in AI use cases, each one comes with a trade-off. Practitioners applying PPML-methods increasingly request an overview of the types and impacts of these trade-offs. To aid this gap in knowledge, this article applies design science research to collect trade-off dimensions and method impacts in an extensive literature review. It then evaluates the specific trade-offs with a focus group of experts and finally constructs an overview over PPML-methods and method combinations’ impact. The final trade-off dimensions are privacy, utility, effort, transparency, and fairness. Seven PPML-methods and their combinations are evaluated according to their impact in these dimensions, resulting in a vast collection of design knowledge and identified research gaps. Keywords: Privacy Preserving Machine Learning, Trade-off, Hybrid Methods, Design Science 1 Introduction Numerous practical domains stand to profit from the capabilities of Artificial Intelligence (AI), raising the heating efficiency in smart living [Ko23], aiding patients and physicians alike in smart health [Na21], and enhancing efficiency in smart audit [ERT23], just to name a few. While the possibilities of AI are virtually unlimited, practical implementation of AI often clashes with a necessity to secure sensitive data [TSG21] . Patient data in healthcare, tenant data in smart living, and company secrets in 1 Strategion GmbH, Albert-Einstein-Straße 1, 49076 Osnabrück, [marian.eleks@strategion.de](marian.eleks@strategion.de), [https://orcid.org/0000-0002-1516-5129](https://orcid.org/0000-0002-1516-5129) 2 Universität Osnabrück, Informationsmanagement und Wirtschaftsinformatik, Hamburger Straße 24, 49084 Osnabrück, [jihler@uni-osnabrueck.de](jihler@uni-osnabrueck.de) 3 Deutsches Forschungszentrum für Künstliche Intelligenz GmbH , Smart Enterprise Engineering, Hamburger Straße 24, 49084 Osnabrück, [jonas.rebstadt@dfki.de](jonas.rebstadt@dfki.de), https://orcid.org/0000-0001-8531-3273 4 Deutsches Forschungszentrum für Künstliche Intelligenz GmbH, Smart Enterprise Engineering, Hamburger Straße 24, 49084 Osnabrück, [henrik.kortum@dfki.de](henrik.kortum@dfki.de), https://orcid.org/0000-0002-1089-711X 5 Universität Osnabrück, Informationsmanagement und Wirtschaftsinformatik , Hamburger Straße 24, 49084 Osnabrück, [oliver.thomas@uni-osnabrueck.de](oliver.thomas@uni-osnabrueck.de) [cba](cba doi:10.18420/inf2024_02)[ doi:10.18420/inf2024_02](cba doi:10.18420/inf2024_02) smart audit are inherently worthy of protection, slowing down the advancement of technologies that depend on these types of data. The field of Privacy Preserving Machine Learning (PPML) presents itself as the natural solution to this conundrum with various methods to preserve the privacy and secrecy of the sensitive data while still enabling machine learning (ML) to be performed. Unfortunately, applying PPML-methods opens up trade-off scenarios with the most prominent being a reduction in output utility [El22]. As AI permeates more and more organizations, people responsible for handling data in these organizations newly adopting AI-solutions are confronted with having to deal with PPML. To be able to efficiently harness the possibilities of PPML, practitioners would greatly benefit from knowing which dimensions of trade-offs to expect and in which way the application of single or hybrid PPML-methods will impact the balance in these dimensions before then diving

FAQs

What trade-off dimensions are identified for privacy preserving machine learning?

The paper concludes five final trade-off dimensions: privacy, utility, effort, transparency, and fairness.

How does the paper study PPML trade-offs and their impacts?

It applies design science research using an extensive literature review to collect trade-off dimensions and method impacts, evaluates specific trade-offs with an expert focus group, and then builds an overview for methods and hybrid combinations.

What are the main PPML method categories discussed in the theoretical foundation?

The paper divides PPML methods into three categories based on where they are applied in the ML process, including data publishing and data processing approaches.

Aug 4, 2026

Upload by Eden