MODELS AND ALGORITHMS FOR POISONING ATTACKS ON MACHINE LEARNING COMPONENTS OF INTRUSION DETECTION SYSTEMS

Ichetovkin Egor

In modern conditions, intrusion detection systems integrating machine learning methods are actively used to detect cyber threats in computer networks. Such solutions demonstrate high efficiency in detecting anomalies, but their reliability can be significantly reduced in the case of poisoning attacks aimed at compromising training. This study provides a detailed analysis of models and algorithms of such attacks, and assesses their impact on the operation of defense mechanisms. The methodological basis of the work includes modeling data poisoning attacks with subsequent assessment of system performance using classical metrics: accuracy, recall, and F-measure. The key novelty of the study lies in its integrated approach. Various machine learning algorithms used to detect anomalies are considered, including: single-class support vector machines, random forest, and deep machine learning. For the first time, comparative testing of several systems under the influence of targeted poisoning attacks was carried out, which made it possible to identify their critical vulnerabilities. The results of the experimental analysis confirmed that the studied systems are susceptible to poisoning attacks, which leads to a significant decrease in their detection ability. In particular, a 15–30% drop in F-score was observed depending on the attack type and the model used. The findings highlight the need to develop more robust learning methods that are resistant to adversarial influences, as well as to modernize existing intrusion detection systems with intelligent classifiers. The work contributes to the development of defense mechanisms by offering not only threat analysis, but also a direction for further research in the field of improving the resilience of machine learning algorithms for intrusion detection systems under poisoning attacks.

Keywords:cybersecurity, intrusion detection systems, poisoning attacks, machine learning component, models and algorithms