Information and communication and chemical technologies

Vol. 2 No. 27 (2025) / 2025-06-28 / Number of views: 248

SIMULATION AND MITIGATION OF DATA POISONING ATTACKS IN MACHINE LEARNING

Authors

Astana IT University
https://orcid.org/0000-0002-5715-4954
Karaganda University of Kazpotrebsouz
https://orcid.org/0000-0001-9273-7402
Astrakhan State Technical University
https://orcid.org/0000-0001-7231-5368
Karaganda University of Kazpotrebsouz
https://orcid.org/0009-0000-3268-867X
Karaganda University of Kazpotrebsouz
https://orcid.org/0000-0002-9677-0266

Keywords

machine learning (ML), data poisoning attacks, availability attack, ensemble learning, out-of-bag (OOB) prediction

Link to DOI:

https://doi.org/10.58805/kazutb.v.2.27-891

How to quote

Amirova, A., et al. “SIMULATION AND MITIGATION OF DATA POISONING ATTACKS IN MACHINE LEARNING”. Vestnik KazUTB, vol. 2, no. 27, June 2025, doi:10.58805/kazutb.v.2.27-891.

Abstract

Machine learning (ML) models are increasingly used in mission-critical applications and are therefore vulnerable to adversarial attacks. Data poisoning attack is one such threat, in which an adversary injects specially designed samples into training data, which can cause a sudden drop in model accuracy or even intentionally bias its decisions in favor of the attacker. This paper presents a simulation of a data poisoning attack on machine learning using MATLAB. Experimental results show that data poisoning can significantly reduce classification accuracy, but using data cleaning techniques can restore model performance. A mitigation method is also proposed that combines the ideas of ensemble learning, outlier prediction estimation, and marginal outlier filtering. Although the main components of the method are widely known, their combination to combat data poisoning attacks is original. Data visualization, decision boundary construction, and accuracy analysis confirm the effectiveness of the mitigation measures. The proposed approach can be useful for developers of secure machine learning systems and researchers studying the resilience of models to adversarial influences. Future research may focus on developing more advanced defense mechanisms such as trust-based learning methods or adaptive attack detection algorithms.