Adversarial examples in Arabic language

Abstract

Adversarial attacks have a great popularity in the artificial intelligence (AI) domain. In the natural language processing (NLP) field, various techniques have been used to evaluate the vulnerability of deep learning (DL) models. It is observed that while most studies focused on generating adversarial examples in English language, Arabic adversarial attacks have received little attention. This paper presents a two-step method to create adversarial examples in Arabic language: first, the most important words are identified. Then, the proposed transformation algorithm is applied. Only small and imperceptible manipulations based on common mistakes in Arabic writing mislead the popular pre-trained language model (PLM) bidirectional encoder representations from transformers (BERT) retrained on the book reviews in Arabic dataset (BRAD) on the sentiment analysis (SA) task and decrease its performance: the classification accuracy was reduced by an average of 3.44%. This drop in accuracy shows that the model was successfully attacked.