Adversarial Examples Generation Method for Chinese Text Classification

被引：0

作者：

Xu, En-Hui ^{[1
]}

Zhang, Xiao-Lin ^{[1
]}

Wang, Yong-Ping ^{[1
]}

Zhang, Shuai ^{[1
]}

Liu, Li-Xin ^{[2
]}

Xu, Li ^{[3
]}

机构：

[1] School of Information Engineering, Inner Mongolia University of Science and Technology, Baotou,014010, China

[2] School of Information, Renmin University of China, Beijing,100872, China

[3] Baotou Medical College, Inner Mongolia University of Science and Technology, Baotou,014010, China

来源：

International Journal of Network Security | 2022年 / 24卷 / 04期

关键词：

Adversarial example - Black boxes - Black-box attack - Chinese character characteristic - Chinese characters - Chinese text classification - Generation method - Morphological features - Phonological features - Text classification systems;

D O I：

10.6633/IJNS.20220724(4).01

中图分类号：

学科分类号：

摘要：

Aiming at the problem that DNNs-based text classification systems are vulnerable to adversarial example attacks, a method of adversarial example generation for Chinese text classification, WordHit, is proposed. In this method, we use the morphological and phonological features of Chinese characters to establish a pool of similar characters and homophones, find important words or phrases that affect classification by removing non-contributing clauses and calculating word importance scores and design a modification strategy that combines word sound and word form to generate adversarial examples to achieve a black-box attack on Chinese text classification models. The word-CNN model and the BiLSTM model are used to verify the effectiveness and versatility of different classification tasks. It is proved that the adversarial example generated by this method can be effectively transferred to the BERT model and the actual deployed sentiment analysis system. © Institute of Mathematical Statistics, 2022

引用

页码：587 / 596