Features of using Markov decision-making processes when modeling attacks on artificial intelligence systems

In this paper, we study the features of modeling attacks on artificial intelligence systems. Markov decision-making processes are used in the construction of the model. A multilevel approach to the interpretation of system states is proposed, which includes several stages of detailing the states. This approach is based on the MITRE ATLAS methodology and the FSTEC Threat Assessment Methodology. When forming the vector, the specifics of the intruder model are taken into account, and two main modeling modes are considered: on-time and off-time. The procedure for the formation of awards at the abstract level (without specifying the actions of the attacker) of building a model is described.