SLAP – simple linear attack for perceptron

A. I. Perminov

Ivannikov Institute for System Programming of the RAS

Abstract: This article introduces a new approach to tricking perceptron based neural networks with piecewise linear activation functions using basic linear algebra. By formulating the attack as a system of linear equations and inequalities, it demonstrates a streamlined and computationally efficient approach to generating diverse sets of adversarial examples. The algorithms for the proposed attack have been implemented in code, that accessible in the open-source repository. The study highlights the formidable challenge posed by the proposed attack methodology for contemporary neural network defenses, emphasizing the pressing need for innovative defense strategies. Through a comprehensive exploration of adversarial vulnerabilities, this research contributes to the advancement of adversarial robustness in machine learning, paving the way for the development of more reliable and trustworthy artificial intelligence systems in real-world applications.

Keywords: neural networks, adversarial attack, linear algebra, trustworthy artificial intelligence

DOI: 10.15514/ISPRAS-2024-36(3)-6