Annotation: The paper is devoted to the problem of ensuring the confidentiality of models in machine learning systems. The aim of the work is to ensure the confidentiality of proprietary models of machine learning systems. In the course of the work we analyzed attacks aimed at violating the confidentiality of models of machine learning systems, as well as ways to protect against this type of attacks, as a result of which the problem of protection against such attacks is set as a search for anomalies in the input data. We propose a way to detect anomalies in the input data based on statistical data, taking into account the resumption of the attack under a different account of the attacker. The obtained results can be used as a basis for designing components of machine learning defense systems.

Keywords: information security, artificial intelligence, artificial intelligence security, attacks on machine learning systems, privacy, model privacy, behavioral analysis.