Abstract
Validation of CNNs is extremely important, especially when they are used in safety-critical domains. In particular, in the
latest years, the focus of validation has been put on assessing the robustness of CNNs, i.e., their ability to correctly classify
perturbed input data. A way to measure robustness is to check the network accuracy over many datasets obtained by altering
the input data in different ways, but this is time and resource-consuming. In this paper, we present ASAP, a method to efficiently
compute the robustness of a CNN, exploiting a parabola-based approximation which allows to adaptively select only relevant
alteration levels. The method is tested on two different benchmarks (MNIST and breast cancer classification). Moreover, we
compare ASAP with other techniques based on uniform sampling, numerical integration, and random sampling.
[download the pdf file] [DOI]