On a formal verification of machine learning systems

The paper deals with the issues of formal verification of machine learning systems. With the growth of the introduction of systems based on machine learning in the so-called critical systems (systems with a very high cost of erroneous decisions and actions), the demand for confirmation of the stability of such systems is growing. How will the built machine learning system perform on data that is different from the set on which it was trained? Is it possible to somehow verify or even prove that the behavior of the system, which was demonstrated on the initial dataset, will always remain so? There are different ways to try to do this. The article provides an overview of existing approaches to formal verification. All the considered approaches already have practical applications, but the main question that remains open is scaling. How applicable are these approaches to modern networks with millions and even billions of parameters? - from our new paper

A Survey of Adversarial Attacks and Defenses for image data on Deep Learning

This article provides a detailed survey of the so-called adversarial attacks and defenses. These are special modifications to the input data of machine learning systems that are designed to cause machine learning systems to work incorrectly. The article discusses traditional approaches when the problem of constructing adversarial examples is considered as an optimization problem - the search for the minimum possible modifications of correlative data that ”deceive” the machine learning system. As tasks (goals) for adversarial attacks, classification systems are almost always considered. This corresponds, in practice, to the so-called critical systems (driverless vehicles, avionics, special applications, etc.). Attacks on such systems are obviously the most dangerous. In general, sensitivity to attacks means the lack of robustness of the machine (deep) learning system. It is robustness problems that are the main obstacle to the introduction of machine learning in the management of critical systems. - from our new paper