Generating Adversarial Examples in Chinese Texts using Mixed-level Perturbations

Compared with the research on adversarial examples on English data, there are few models for generating adversarial examples on Chinese data. Most of the Chinese adversarial examples are in a single form. Their fluency and attack accuracy are not well performed. In this paper, we propose MixAttacker that uses word-level and sentence-level perturbations in conjunction with each other to generate adversarial examples in Chinese texts. The model uses the masked language model WoBERT [1] to generate replacement words based on the Chinese word-level transformation, then selects one of the new sentences obtained by word replacement by merit, and finally, back translate this sentence. In addition, we propose sentence fluency evaluation to control the quality of adversarial examples more effectively. The experimental results show that our model achieves 75.50%, 69.00%, 46.00%, and 55.50% accuracy decrease on four datasets, respectively, ChnsentiCorp, Hotel, TUHCNews, and Weibo, with effective perturbation, semantic and fluency control.