PROTECTION AGAINST ATTACKS ON MACHINE LEARNING SYSTEMS ON THE EXAMPLE OF EVADIATION ATTACKS IN MEDICAL IMAGE ANALYSIS
E. A. Rudnitskaya, M. A. Poltavtseva
Peter the Great St.Petersburg Polytechnic University
Annotation: This paper is about the adversarial attacks on machine learning systems that analyze medical images. The authors review the existing attacks, conducts their systematization and practical feasibility. The article contains an analysis of existing methods of protection against adversarial attacks on machine learning systems. It describes the peculiarities of medical images. The authors solve the problem of protection against adversarial attacks for these images based on several defensive methods. The authors have determined the most relevant protection methods, their implementation and testing on practical examples – the analysis of COVID-19 patient’s images.
Keywords: attacks on machine learning systems, machine learning system protection, adversarial attacks, medical images, machine learning