Bigram and Unigram Based Text Attack via Adaptive Monotonic Heuristic Search

Deep neural networks (DNNs) are known to be vulnerable adversarial images, while their robustness in text classification rarely studied. Several lines of attack methods have been proposed the literature, such as character-level, word-level, and sentence-level attacks. However, it is still a challenge minimize number word distortions necessary induce misclassification, simultaneously ensuring lexical correctness, syntactic semantic similarity. In this paper, we propose Bigram Unigram based Monotonic Heuristic Search (BU-MHS) method examine vulnerability deep models. Our has three major merits. Firstly, documents not only at unigram level but also bigram avoid producing meaningless outputs. Secondly, hybrid replace input words with both synonyms sememe candidates, which greatly enriches potential substitutions compared using synonyms. Lastly, design search algorithm, i.e., (MHS), determine priority replacements, aiming reduce modification cost an attack. We evaluate effectiveness BU-MHS on IMDB, AG's News, Yahoo! Answers datasets by attacking four state-of-the-art DNNs Experimental results show that our achieves highest success rate changing smallest other existing