Modeling scenarios of destructive impact on the integrity of machine learning models

A. B. Menisov, A. G. Lomako

Mozhaiskiy Space Military Academy

Abstract: The article is devoted to the development of models of destructive impact on the integrity of machine learning models based on SIR forecasting of the scale of threats and risks of losses under various scenarios of computer attacks. The article presents an original model of information security threats to technical components of artificial intelligence in the context of heterogeneous mass computer attacks, displaying vulnerabilities and methods of possible enemy actions. The authors have developed a methodology for adapting modernized SIR models of natural epidemics to identify similarities and analogues in the nature of the spread of destructive failures in AI systems caused by heterogeneous mass and targeted impacts. The identified patterns made it possible to assess the risks of possible damage to integrity and develop effective strategies for preventing and correcting distortions of machine learning models.

Keywords: artificial intelligence, integrity of machine learning models, trust, information security, diagnostic testing, test environment.

DOI: 10.15514/ISPRAS-2025-37(3)-4