Abstract
Several adversarial attacks have been pro-posed in the domains of computer vision and natural language processing (NLP). However, most attacks in the NLP domain have been applied to evaluate deep neural networks (DNNs) that were trained on English corpora. This paper proposes the first set of character-level adversarial attacks designed for models trained on Arabic. We present an efficient method to generate character-level adversarial examples against neural classifiers. Our method relies on flip operations that were designed based on the most common spelling mistakes that non-native Arabic learners make. We find that only a few manipulations are needed to mislead powerful and popular DNN-based classifiers trained on Arabic corpora.