Textual Adversarial Attacking with Limited Queries

Recent studies have shown that natural language processing (NLP) models are vulnerable to adversarial examples, which maliciously designed by adding small perturbations benign inputs imperceptible the human eye, leading false predictions target model. Compared character- and sentence-level textual attacks, word-level attack can generate higher-quality especially in a black-box setting. However, existing methods usually require huge number of queries successfully deceive model, is costly real scenario. Hence, finding appropriate difficult. Therefore, we propose novel method, main idea fully utilize examples generated local model transfer part complete ahead time, thereby reducing costs related attacking Extensive experiments conducted on three public benchmarks show our method not only improve success rate but also reduce cost, while outperforming baselines significant margin.