Effective inclusion methods for verification of ReLU neural networks

Szász, Attila and Bánhelyi, Balázs (2024) Effective inclusion methods for verification of ReLU neural networks. ANNALES MATHEMATICAE ET INFORMATICAE, 60. pp. 141-150. ISSN 1787-6117

Preview

Text AMI_60_from141to150.pdf - Published Version Download (538kB) | Preview

Abstract

The latest machine learning models are sensitive to adversarial inputs, i.e., the neural network can give incorrect results even with small changes in the learning case. To avoid this, techniques are used during learning, or verification is also possible. In many cases, these methods use interval arithmetic, whose usefulness is severely limited by overestimation. In this paper, we present and compare such methods that can handle this problem.

Actions (login required)

Edit Item