Camouflage as adversarial attacks on machine learning models

Abstract

The article is devoted to adversarial attacks on machine learning models. Such attacks are understood as the deliberate manipulation of data at different stages of the machine learning pipeline, designed to either prevent the correct operation of the model, or to achieve the desired result from it. In this case, physical evasion attacks were considered, that is, the objects themselves used in the model were modified. The article discusses the use of camouflaged images to deceive the recognition system. The experiments were carried out with a machine learning model that recognizes images of cars. Two types of camouflage patterns were used - classic camouflage and drawing an image of another car. In practice, such manipulations can be carried out using airbrushing. The work confirmed the successful implementation of such attacks, and obtained metrics characterizing the effectiveness of attacks, as well as the possibility of countering them using competitive training. All results are openly published, which makes it possible to use them as a software stand for testing other attacks and protection methods.