On improving the robustness of machine learning models

Abstract

This article discusses how to improve robustness for machine learning models. Robustness is one of the most important characteristics of machine learning models, which determines the possibility of practical use of the models. However, not everything is simple when determining this characteristic for specific models. Firstly, not everything is clear with the very definition of robustness. If we consider robustness as the preservation of the behavior of the model under small perturbations of the initial data, then at least two questions arise - how small should these changes be, and how does such a definition relate to other characteristics of the model? First of all, among other characteristics, it is necessary to consider the generalizing ability of the model (generalization), which is determined by the model’s work with previously unknown data. Google's concept of model reliability is also discussed. The main content of the article is devoted to the consideration of competitive training, which, despite all its limitations, remains today the main tool for increasing robustness.