多约束引导的中文对抗样本生成  

Chinese Adversarial Example Generation Guided by Multi-constraints

作　　者：韩子屹 王巍[1] 玄世昌[1] HAN Ziyi;WANG Wei;XUAN Shichang(College of Computer Science and Technology,Harbin,Heilongjiang 150001,China)

机构地区：[1]哈尔滨工程大学计算机科学与技术学院,黑龙江哈尔滨150001

出　　处：《中文信息学报》2023年第2期41-52,共12页Journal of Chinese Information Processing

基　　金：NSFC-新疆联合基金(U2003206);国家自然科学基金(61972255);中央高校基本科研业务费专项资金(GK2060260303);黑龙江省重点研发计划项目(GA21C020)。

摘　　要：深度神经网络(DNN)已经被广泛应用于图像识别和自然语言处理等各个领域。近年来的研究表明,向DNN模型输入包含微小扰动的样本后,很容易对其输出结果造成严重破坏,这样处理过的样本被称为对抗样本。但中文对抗样本生成领域一直面临着一个严重问题,攻击成功率和对抗样本的可读性难以兼得。该文提出了一种在对抗样本生成的不同阶段,对对抗样本进行视觉相似度和语义相似度进行约束的对抗攻击方法MCGC。MCGC生成的对抗样本不但具有良好的可读性,且在针对Text-CNN、Bi-LSTM、BERT-Chinese等多个模型的定向和非定向攻击可以达到90%左右的攻击成功率。同时,该文还研究了以BERT-Chinese为代表的掩码语言模型(MLM)和传统自然语言处理模型在鲁棒性上的差异表现。Recent studies have shown that feeding DNNs with adversarial samples,i.e.,samples containing small perturbations,can easily wreak havoc on their output.The field of Chinese adversarial sample generation has been challenged by achieving both the attack success rate and the sample readability.In this paper,we propose an adversarial attack method named MCGC that constrains the visual similarity and semantic similarity of adversarial samples at different stages of adversarial sample generation.Such generated adversarial samples have good readability and achieve an 90%or so success rate in target and untarget attacks against multiple models such as Text-CNN,Bi-LSTM,and BERT-Chinese.At the same time,this paper studies the differences in robustness between the mask language models(MLM)represented by BERT and the traditional natural language processing models.

关 键 词：字形相似度评估 语义相似度控制 黑盒对抗攻击 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]