Machine learning security assessment method based on adversary and attack methods

Analytical methods for assessing the security of Machine Learning Systems (MLS) that have been proposed in other researches do not provide compatibility with each other and their taxonomies have become incomplete due to the introduction of new properties of adversarial machine learning. In this sense, we have identified carefully relevant concepts of most prevalent researches about the security assessment of MLS. We propose a novel security assessment method based on the modeling of the adversary and the selection of adversarial attack methods for the generation of adversarial examples related to the also proposed taxonomy. This method provides compatibility with other proposed methods as well as practical guidelines and tools for evaluating machine learning systems. We also introduce the concern for efficient metrics capable of measuring the robustness of MLS to adversarial examples. This research is focused on the empirical evaluation of the security of machine learning systems, rather than on classical performance evaluation.