基于多变量时序数据的对抗攻击与防御方法  

Adversarial Attack and Defense Method Based onMultivariable Time Series Data

作　　者：刘坤 曾恩 刘博涵 李俊达 李江荣[2] LIU Kun;ZENG En;LIU Bohan;LI Junda;LI Jiangrong(School of Automation,Bejing Institute of Technology,Bejing 100081,China;College of Mathematics and Computer Science,Yan'an University,Yan'an 7160,China)

机构地区：[1]北京理工大学自动化学院,北京100081 [2]延安大学数学与计算机科学学院,延安716000

出　　处：《北京工业大学学报》2023年第4期415-423,共9页Journal of Beijing University of Technology

基　　金：国家自然科学基金资助项目(62273041,61873034);流程工业综合自动化国家重点实验室联合开放基金资助项目(2021-KF-21-05)。

摘　　要：为了保证时序数据攻击检测模型的安全性,提出了基于多变量时间序列数据的对抗攻击与对抗防御方法.首先,针对基于自编码器的攻击检测模型,设计了在测试阶段实施的逃逸攻击.其次,针对设计的对抗攻击样本,提出了一种基于雅克比(Jacobian)正则化方法的对抗攻击防御策略,将模型训练过程中的Jacobian矩阵作为目标函数中的正则项,提高了深度学习模型对对抗攻击的防御能力.在工业水处理数据集BATADAL上进行实验,验证了所提出的对抗攻击方法和对抗防御方法的有效性.To ensure the security of the attack detection model of time series data,an adversarial atack and adversarial defense method based on multivariate time series data was proposed.First,the escape attack implemented in the test phase was designed for the autoencoder-based attack detection model.Second,according to the designed adversarial attack samples,the adversarial defense strategy based on the Jacobian regularization method was proposed.The Jacobian matrix in the calculation model training process was taken as the regular term in the objective function to improve the defense capability of the deep learning model.The attack effects of the proposed attack methods and the defense effect of the proposed adversarial defense method were verified on the BATADAL dataset of industrial water treatment.

关 键 词：多变量时序数据 攻击检测 对抗攻击 对抗防御 自编码器 Jacobian正则化 

分 类 号：U461[机械工程—车辆工程] TP308[交通运输工程—载运工具运用工程]