面向时序数据的多范数约束对抗样本生成方法  

Generation Method of Adversarial Samples with Multi-norm Constraints for Time Series Data

作　　者：张鑫 沈子钰 李云[1] ZHANG Xin;SHEN Ziyu;LI Yun(School of Computer Science,Nanjing University of Posts and Telecommunications,Nanjing Jiangsu 210023,China)

机构地区：[1]南京邮电大学计算机学院,江苏南京210023

出　　处：《指挥与控制学报》2023年第3期253-262,共10页Journal of Command and Control

基　　金：国家自然科学基金(61772284)资助。

摘　　要：云控制系统中的传感器产生了大量的时序数据,需要对这些数据进行分类以作出自动化决策.随着人工智能算法在云控制系统中的广泛应用,其安全性问题显得尤为重要.为了研究云控制系统中深度模型的安全性问题,提出一种新的针对深度时序模型的对抗攻击算法,定义一个新的时序数据对抗样本性能指标,使用UCR数据集进行实验,验证该算法的优异性能,展现云控制系统中深度模型的脆弱性.Sensors in cloud control systems generate a large amount of time series data,which need to be classified to make automated decisions.As artificial intelligence algorithms are widely applied in cloud control systems,their security issues are particularly important.To study the security problems of deep models in cloud control systems,a new adversarial attack algorithm for deep time series models is proposed,and a new performance index for adversarial samples of time series data is defined.The UCR dataset is used to conduct experiments,the superior performance of the algorithm is verified and the vulnerability of deep models in cloud control systems is shown.

关 键 词：云控制系统 深度学习 时间序列分类 对抗攻击 

分 类 号：TP311.13[自动化与计算机技术—计算机软件与理论] TP18[自动化与计算机技术—计算机科学与技术]