自然语言处理中对抗攻防的可解释性综述  

A review of the interpretability of the confrontation and defenses in natural language processing

作　　者：胡瑞斌 苏世帅 王哲 HU Ruibin;SU Shishuai;WANG Zhe(Department of Science and Technology,Zhejiang Normal University,Jinhua 321004,China;School of Computer Science and Technology,Zhejiang Normal University,Jinhua 321004,China)

机构地区：[1]浙江师范大学科技处,浙江金华321004 [2]浙江师范大学计算机科学与技术学院,浙江金华321004

出　　处：《浙江师范大学学报(自然科学版)》2025年第2期142-153,共12页Journal of Zhejiang Normal University:Natural Sciences

基　　金：浙江省教育厅一般科研项目(Y202457289);浙江省教育厅理工类一般科研项目(Y202456821)。

摘　　要：针对深度学习模型安全性面临的巨大挑战,特别是易受对抗样本攻击的问题,整理近年来对抗样本与模型可解释性方面的研究进展.通过系统梳理对抗攻击和防御相关的可解释性研究,分析对抗样本的生成方法及其对模型决策过程的影响,并讨论防御对抗攻击时采用的可解释性技术和提升模型鲁棒性的方法.重点阐述了近年来对抗样本对模型分类结果的影响,可解释性技术在揭示模型决策过程脆弱性方面的有效性,以及结合可解释性的防御方法在提升模型鲁棒性方面的进展.为可解释性研究提供有价值的参考,推动深度神经网络模型在自然语言处理等领域的安全性研究.In response to the huge challenges facing deep learning models,especially the problem of being vulnerable to adversarial sample attacks,it was sorted out the research progress in adversarial samples and model interpretability in recent years.Through a systematic analysis of interpretability research related to adversarial attacks and defenses,it was examined the methods for generating adversarial examples and their impact on model decision-making.It was also discussed the interpretability techniques employed in defending against adversarial attacks and methods for improving model robustness.Focusing on recent advancements,it was highlighted the impact of adversarial examples on model classification results,the effectiveness of interpretability techniques in revealing vulnerabilities in model decision-making processes,and the progress achieved by combining interpretability with defense methods to enhance model robustness.The research was also aimed to provide valuable insights for interpretability research and promote the development of more secure deep neural network models in natural language processing and other fields.

关 键 词：深度神经网络 自然语言处理 可解释性 对抗攻击 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]