INTELLIGENT SYSTEMS AND TECHNOLOGIES

Detecting trigger!type attacks on artificial neural networks

A. A. Zatsarinnyi^a, A. A. Karandeev^b, B. N. Chetverushkin^b, V. P. Osipov^b, N. A. Yashin<sup>b

a</sup> Federal Research Center "Computer Science and Control" of Russian Academy of Sciences, Moscow, Russia
^b Keldysh Institute of Applied Mathematics of Russian Academy of Sciences, Moscow, Russia

Abstract: The article addresses the pressing issue of trigger-based attacks on artificial neural networks designed for image recognition in the context of ensuring their reliability and security. Various scenarios of trigger-based attacks, their main implementation methods, and the consequences of such attacks are examined. The article provides a detailed analysis of methods for applying triggers to images, approaches for detecting triggers, including identifying key characteristics inherent to images containing triggers. The results of the proposed method for combating trigger-based attacks are presented, enabling the detection of triggers in images during the machine learning phase of neural networks. The prospects for developing protection methods against trigger-based attacks in the context of machine learning and convolutional neural networks are also discussed.

Keywords: artificial neural network, machine learning, trigger-based attack, image recognition.

DOI: 10.14357/20718632250101

Bibliographic databases:

•