Resumen
With the growing use of systems based on machine learning, which, from a practical point of view, are considered as systems of artificial intelligence today, attention to the issues of reliability (stability) of such systems and solutions is also growing. For so-called critical applications such as real-time decision-making systems, special systems, etc. sustainability issues are crucial from the point of view of the practical use of machine learning systems. The use of machine learning systems (artificial intelligence systems, which is now, in fact, a synonym) in such areas is possible only with the proof of stability (determination of guaranteed performance parameters). Resiliency problems arise from different characteristics of the data during training (training) and testing (practical application). At the same time, additional complexity is created by the fact that, in addition to natural reasons (unbalanced samples, measurement errors, etc.), the data can be deliberately modified. These are the so-called attacks on machine learning systems. Accordingly, it is impossible to talk about the reliability of machine learning systems without protection against such actions. In this case, attacks can be directed both at the data and at the models themselves.