One of the important elements in evaluating the stability of machine learning systems are the so-called adversarial examples. These are specially selected or artificially created input data for machine learning systems that interfere with their normal operation, are interpreted or processed incorrectly. Most often, such data are obtained through some formal modifications of the real source data. This article considers a different approach to creating such data, which takes into account the semantic significance (meaning) of the modified data - counterfactual examples. The purpose of the work is to present practical solutions for generating counterfeit examples. The consideration is based on the real use of counterfactual examples in assessing the robustness of machine learning systems.
No comments:
Post a Comment