Characteristic Analysis of Methods to Counteract Adversarial Machine Learning Attacks

In the era of technological evolution and digital transformation, machine learning and artificial intelligence take a significant part of intelligent systems’ development. Together with the integration of smart and autonomous operations in everyday life, novel security issues arise. One such a malicious threat against the security of machine learning models is the adversarial machine learning (AML) attack. It consists in alteration of the original data samples used to train or test the model with indistinguishably small feature modifications that results in inaccurate processing outcome. In this paper, the most significant strategies to counteract AML attacks are studied and compared through several characteristics that have main influence on their successful implementation in different realization scenarios.