Impact of k-anonymity on machine learning model performance: an empirical study
Scientific Bulletin of UNFU · 25.06.2026 · 36 (3)
DOI: 10.36930/40360310 ↗Kurz gefasst
Fachartikel in der Zeitschrift Scientific Bulletin of UNFU, online erschienen am 25.06.2026. Verfasst von А. М. Вахула und 1 weiteren Autor:innen. Die Arbeit nennt 20 Literaturangaben.
Abstract (Originalsprache)
The rapid growth of digital data and the widespread use of machine learning techniques have increased concerns regarding the protection of personal information. Data anonymization is widely used to mitigate privacy risks by transforming datasets before analysis. However, such transformations may reduce data utility and negatively affect the performance of machine learning models. This study investigates the impact of k-anonymity-based data anonymization on classification performance using the Adult Income dataset. In the dataset, personal data were marked as quasi-identifiers. Quasi-identifying attributes are generalized and suppressed to achieve different levels of anonymity (k = 5, 10, 20), and machine learning models are trained on both original and anonymized datasets. After training, models are evaluated using a subset of the original dataset. Performance of each k-anonymity level in the experiment is based on accuracy, recall, precision, and F1-score. The experimental results demonstrate that increasing the level of anonymization leads to a gradual decrease in predictive performance. Accuracy is only moderately affected: it decreases from 0.828-0.84 on the original dataset to 0.798-0.81 on strongly anonymized data. A more significant reduction is observed in recall and F1-score: from 0.586-0.626 on the original dataset to 0.435-0.555 on strongly anonymized data. Those changes indicate a loss in the ability of models to correctly identify positive instances, which harm performance and reliability of the model. At the same time, moderate anonymization levels, in the range k = 5-10, provide a reasonable balance between privacy protection and analytical utility. The findings confirm the existence of a trade-off between privacy preservation and model performance and highlight the importance of selecting appropriate anonymization parameters in machine learning workflows. Additionally, some machine learning models, like Gradient Boosting, have shown better performance on anonymized data than others, which may help to select an appropriate algorithm for a specific level of anonymized data.
Zitieren
APA
Вахула, А. М., & Іванюк, О. О. (2026). Impact of k-anonymity on machine learning model performance: an empirical study. Scientific Bulletin of UNFU, 36(3). https://doi.org/10.36930/40360310
Vancouver
Вахула АМ, Іванюк ОО. Impact of k-anonymity on machine learning model performance: an empirical study. Scientific Bulletin of UNFU. 2026;36(3). doi:10.36930/40360310
BibTeX
@article{2026,
author = {Вахула, А. М. and Іванюк, О. О.},
title = {Impact of k-anonymity on machine learning model performance: an empirical study},
journal = {Scientific Bulletin of UNFU},
year = {2026},
volume = {36},
number = {3},
doi = {10.36930/40360310}
}
EUVIMED