面向深度神经网络的后门攻击研究综述  

Overview of the Research on Backdoor Attacks for DNN

作　　者：易月娥[1] 程玉柱[1] YI Yuee;CHENG Yuzhu(Changsha Social Work College,Changsha,Hunan,China 410004)

机构地区：[1]长沙民政职业技术学院,湖南长沙410004

出　　处：《湖南邮电职业技术学院学报》2023年第3期37-41,共5页Journal of Hunan Post and Telecommunication College

基　　金：2022年湖南省自然科学基金项目“面向新一代互联网的高性能包分类关键技术研究”(课题编号:2022JJ60099)。

摘　　要：随着深度神经网络(deep neural networks,DNN)的广泛应用,深度神经网络模型的安全性问题日益突出。后门攻击是一种常见的攻击方式,其目的在于恶意改变DNN模型训练后的表现而不被人类视觉发现。文章介绍了深度神经网络及其后门攻击的概念,详细描述了深度学习模型中的后门攻击范式、后门评估指标、后门设置及计算机视觉领域的后门攻击等内容,并对其优缺点进行了总结和评析,此外还介绍了后门攻击技术在相关领域的一些积极应用。最后,对未来DNN后门防御技术研究进行了展望。With the widespread application of deep neural networks(DNN),the security issues of deep neural network models are becoming increasingly prominent.Backdoor attack is a common attack method,with the aim of maliciously altering the performance of DNN model training without being detected by human vision.The article introduces the concepts of deep neural networks and their backdoor attacks,and provides a detailed description of the backdoor attack paradigm,backdoor evaluation indicators,backdoor settings,and backdoor attacks in the field of computer vision in deep learning models.It also summarizes and evaluates their advantages and disadvantages.In addition,it introduces some positive applications of backdoor attack technology in related fields.Finally,the outlook and development direction for future research on DNN backdoor defense technology are proposed.

关 键 词：深度神经网络 后门攻击 防御技术 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]