Generating Natural Language Attacks in a Hard Label Black Box Setting

We study an important and challenging task of attacking natural language processing models in a hard label black box setting. propose decision-based attack strategy that crafts high quality adversarial examples on text classification entailment tasks. Our proposed leverages population-based optimization algorithm to craft plausible semantically similar by observing only the top predicted target model. At each iteration, procedure allow word replacements maximizes overall semantic similarity between original text. Further, our approach does not rely using substitute or any kind training data. demonstrate efficacy through extensive experimentation ablation studies five state-of-the-art across seven benchmark datasets. In comparison attacks prior literature, we are able achieve higher success rate with lower perturbation percentage too highly restricted