ML辅助网络自动化系统的对抗样本攻击方法  

An Adversarial Sample Attack for ML-Assisted Network Automation

作　　者：潘小琴 尹慧 蔡熠 段康容 PAN Xiaoqin;YIN Hui;CAI Yi;DUAN Kangrong(Center of Engineering and Technology,Southwest University of Science and Technology,Mianyang 621010;No.722 Research Institute of China State Shipbuilding Corporation Limited,Wuhan 430205)

机构地区：[1]西南科技大学工程技术中心,绵阳621010 [2]中国船舶集团有限公司第七二二研究所,武汉430205

出　　处：《计算机与数字工程》2023年第12期2821-2826,2858,共7页Computer & Digital Engineering

基　　金：西南科技大学博士基金项目(编号:23zx7123)资助。

摘　　要：针对网络自动化系统中用于异常检测的基于机器学习(Machine Learning,ML)的分类器,提出了一种黑盒对抗样本攻击方法来误导分类器输出错误的分类结果。首先,设计了一种对抗样本生成算法来人工合成替代分类器的训练数据集,算法不仅基于一组仅包含“正常”类型的合法遥测数据生成涵盖所有异常类型的合成数据,而且标记数据时还最小化对目标分类器的查询次数。然后,利用对抗样本去攻击了系统中基于ML的分类器,分析了对抗样本给分类器性能带来的影响,并将结果推广到不同结构的ML模型。最后,利用从真实的IP-over-EON多层网络测试平台采集的遥测数据进行了仿真实验,实验结果验证了该方法的有效性。This paper proposes a black-box adversarial attack scheme to mislead the machine learning(ML)based classifiers for anomaly detection to output incorrect classification results in network automation.Firstly,an adversarial sample generation algo-rithm is designed to synthesize training data for the substitute classifier.The algorithm generates synthetic data to cover all the anom-aly types based on a set of legitimate telemetry data that only containing the"normal"type,and label the synthesized data with mini-mized queries to the legitimate classifier.Then,the generated adversarial sample is leveraged to attack the ML-based classifier to analyze the influence of the adversarial samples on performance of the model,and extend the results to different types of models.Fi-nally,extensive simulations are conducted with the telemetry data collected from a real-world IP over elastic optical network(IP-over-EON)testbed.The results show the effectiveness of the proposed scheme.

关 键 词：网络自动化 黑盒攻击 机器学习 异常检测 对抗样本 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]