Efficient Combinatorial Optimization for Word-Level Adversarial Textual Attack

Over the past few years, various word-level textual attack approaches have been proposed to reveal vulnerability of deep neural networks used in natural language processing. Typically, these involve an important optimization step determine which substitute be for each word original input. However, current research on this is still rather limited, from perspectives both problem-understanding and problem-solving. In paper, we address issues by uncovering theoretical properties problem proposing efficient local search algorithm (LS) solve it. We establish first provable approximation guarantee solving general cases. Extensive experiments involving 5 NLP tasks, 8 datasets 26 models show that LS can largely reduce number queries usually order magnitude achieve high success rates. Further adversarial examples crafted higher quality, exhibit better transferability, bring more robustness improvement victim training.