Analysis of an Existing Method for Detecting Adversarial Attacks on Deep Neural Networks

Abstract

Analyzes the existing method of detecting adversarial attacks on deep neural networks, proposed by researchers from Carnegie Mellon University and the Korean Institute of Advanced Technologies (KAIST) Ko, G. and Lim, G in 2021. Examines adversarial attacks, as well as the history of research on the topic. The paper considers the concepts of interpreted and not interpreted neural networks and features of methods of protection of the types of neural networks considered. The method for protecting against adversarial attacks is also considered to be applicable to both types of neural networks. An example of an attack simulation is given, which makes it possible to identify a sign showing that an attack has been committed.