Adversarial Turing Patterns from Cellular Automata

State-of-the-art deep classifiers are intriguingly vulnerable to universal adversarial perturbations: single disturbances of small magnitude that lead misclassification most inputs. This phenomena may potentially result in a serious security problem. Despite the extensive research this area, there is lack theoretical understanding structure these perturbations. In image domain, certain visual similarity between patterns, represent perturbations, and classical Turing which appear as solution non-linear partial differential equations underlying concept many processes nature. paper, we provide bridge two different theories, by mapping simplified algorithm for crafting perturbations (inhomogeneous) cellular automata, latter known generate patterns. Furthermore, propose use generated experimentally show they significantly degrade performance learning models. We found method be fast efficient way create data-agnostic quasi-imperceptible perturbation black-box scenario. The source code available at https://github.com/NurislamT/advTuring.