Anonymisation Versus Encryption: A Comprehensive Evaluation of Privacy-Enhancing Technologies for Machine Learning
The increasing use of open data for training machine learning models creates a tension between data utility and data privacy, particularly under the strict requirements of the General Data Protection Regulation (GDPR). Traditional anonymisation techniques aim to reduce the risk of re-identification but often impair model performance. Homomorphic encryption (HE), by contrast, allows encrypted computation while preserving data fidelity, but introduces significant computational demands. This paper presents a comparative study of anonymisation and HE using three public tabular datasets: Adult, Titanic, and OULA. Results show that HE can preserve or even improve model accuracy, with a maximum gain of 4.9 % in the Titanic dataset and a loss of only 2.3 % in the Adult dataset. However, it increases inference time by up to 9000x and storage by up to 300x. Anonymisation maintains stable accuracy with negligible performance and memory overhead. A legal analysis confirms that HE qualifies as pseudonymisation under the GDPR, offering clear benefits for compliant and secure data sharing. These findings provide practical guidance for selecting privacy-enhancing technologies in machine learning applications, taking into account legal, technical, and operational factors.
You may also find the article on the publisher's website.
Keywords: Anonymisation; privacy; re-identification risk; GDPR; open data; homomorphic encryption; privacy enhancing technologies
Year: 2026
Full text [628 kB]Authors of this publication:
Dalibor Fiala
Phone: +420 377 63 2429
E-mail: dalfia@kiv.zcu.cz
WWW: http://www.kiv.zcu.cz/~dalfia/
Michael Heigl
E-mail: heigl@kiv.zcu.cz