Word-level textual adversarial attacking based on genetic algorithm
Dingmeng Shi, Zhaocheng Ge, Tengfei Zhao · 2022
Since the discovery of adversarial examples, the research on adversarial examples in the image field has caused an academic boom. In recent years, with the development of artificial intelligence, adversarial samples in the text field have also attracted more and more scholars' research interest. This paper proposes such an adversarial sample generation algorithm in a black box scenario: using a targeted word deletion scoring mechanism, it can find keywords that have a significant impact on the decision of the classifier when the internal structure of the model is unknown, and use the HowNet vocabulary to search the synonyms of these keywords are replaced to generate a set of adversarial samples that are semantically consistent with the original samples. Then combined with genetic algorithm to search for the best sample in the generated sample space. The results of testing LSTM and CNN on sentiment classification and news classification data sets show that the algorithm can greatly reduce the accuracy of the target model with less disturbance.