ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2021, Vol. 58 ›› Issue (8): 1727-1750.doi: 10.7544/issn1000-1239.2021.20210304

所属专题: 2021人工智能前沿进展专题

• 人工智能 • 上一篇    下一篇

面向自然语言处理的对抗攻防与鲁棒性分析综述

郑海斌1,陈晋音1,2,章燕1,张旭鸿3,葛春鹏4,刘哲4,欧阳亦可5,纪守领6   

  1. 1(浙江工业大学信息工程学院 杭州 310023);2(浙江工业大学网络空间安全研究院 杭州 310023);3(浙江大学控制科学与工程学院 杭州 310063);4(南京航空航天大学计算机科学与技术学院 南京 211106);5(华为技术有限公司南京研究所 南京 210029);6(浙江大学计算机科学与技术学院 杭州 310063) (haibinzheng320@gmail.com)
  • 出版日期: 2021-08-01
  • 基金资助: 
    国家自然科学基金项目(62072406);浙江省自然科学基金项目(LY19F020025);宁波市“科技创新2025”重大专项(2018B10063)

Survey of Adversarial Attack, Defense and Robustness Analysis for Natural Language Processing

Zheng Haibin1, Chen Jinyin1,2, Zhang Yan1, Zhang Xuhong3, Ge Chunpeng4, Liu Zhe4, Ouyang Yike5, Ji Shouling6   

  1. 1(College of Information Engineering, Zhejiang University of Technology, Hangzhou 310023);2(Cyberspace Security Research Institute, Zhejiang University of Technology, Hangzhou 310023);3(College of Control Science and Engineering, Zhejiang University, Hangzhou 310063);4(College of Computer Science and Technology, Nanjing University of Aeronautics and Astronautics, Nanjing 211106);5(Nanjing Research Center, Huawei Technologies Co., Ltd., Nanjing 210029);6(College of Computer Science and Technology, Zhejiang University, Hangzhou 310063)
  • Online: 2021-08-01
  • Supported by: 
    This work was supported by the National Natural Science Foundation of China (62072406), the Natural Science Foundation of Zhejiang Province (LY19F020025), and the Major Special Funding for “Science and Technology Innovation 2025” in Ningbo (2018B10063).

摘要: 随着人工智能技术的飞速发展,深度神经网络在计算机视觉、信号分析和自然语言处理等领域中都得到了广泛应用.自然语言处理通过语法分析、语义分析、篇章理解等功能帮助机器处理、理解及运用人类语言.但是,已有研究表明深度神经网络容易受到对抗文本的攻击,通过产生不可察觉的扰动添加到正常文本中,就能使自然语言处理模型预测错误.为了提高模型的鲁棒安全性,近年来也出现了防御相关的研究工作.针对已有的研究,全面地介绍自然语言处理攻防领域的相关工作,具体而言,首先介绍了自然语言处理的主要任务与相关方法;其次,根据攻击和防御机制对自然语言处理的攻击方法和防御方法进行分类介绍;然后,进一步分析自然语言处理模型的可验证鲁棒性和评估基准数据集,并提供自然语言处理应用平台和工具包的详细介绍;最后总结面向自然语言处理的攻防安全领域在未来的研究发展方向.

关键词: 深度神经网络, 自然语言处理, 对抗攻击, 防御, 鲁棒性

Abstract: With the rapid development of artificial intelligence, deep neural networks have been widely applied in the fields of computer vision, signal analysis, and natural language processing. It helps machines process understand and use human language through functions such as syntax analysis, semantic analysis, and text comprehension. However, existing studies have shown that deep models are vulnerable to the attacks from adversarial texts. Adding imperceptible adversarial perturbations to normal texts, natural language processing models can make wrong predictions. To improve the robustness of the natural language processing model, defense-related researches have also developed in recent years. Based on the existing researches, we comprehensively detail related works in the field of adversarial attacks, defenses, and robustness analysis in natural language processing tasks. Specifically, we first introduce the research tasks and related natural language processing models. Then, attack and defense approaches are stated separately. The certified robustness analysis and benchmark datasets of natural language processing models are further investigated and a detailed introduction of natural language processing application platforms and toolkits is provided. Finally, we summarize the development direction of research on attacks and defenses in the future.

Key words: deep neural network, natural language processing, adversarial attack, defense, robustness

中图分类号: