汽车工程 ›› 2024, Vol. 46 ›› Issue (8): 1394-1402.doi: 10.19562/j.chinasae.qcgc.2024.08.006

• • 上一篇    

基于BEGAN的汽车CAN网络入侵检测数据增强方法研究

汪想1,刘蓬勃1,赵剑1(),范科峰2,李琳辉1   

  1. 1.大连理工大学汽车工程学院,大连 116024
    2.中国电子技术标准化研究院,北京 100007
  • 收稿日期:2023-08-01 修回日期:2023-10-13 出版日期:2024-08-25 发布日期:2024-08-23
  • 通讯作者: 赵剑 E-mail:jzhao@dlut.edu.cn
  • 基金资助:
    国家自然科学基金联合基金项目(U1930206)

Research on Data Enhancement Methods of BEGAN-Based Intrusion Detection in Automotive CAN Networks

Xiang Wang1,Pengbo Liu1,Jian Zhao1(),Kefeng Fan2,Linhui Li1   

  1. 1.School of Automotive Engineering,Dalian University of Technology,Dalian  116024
    2.China Electronics Standardization Institute,Beijing  100007
  • Received:2023-08-01 Revised:2023-10-13 Online:2024-08-25 Published:2024-08-23
  • Contact: Jian Zhao E-mail:jzhao@dlut.edu.cn

摘要:

针对目前汽车CAN网络入侵检测算法因攻击样本缺少而导致数据不平衡问题,提出一种基于BEGAN的CAN入侵检测数据增强方法,引入one-hot编码将CAN报文特征图像化,结合构建的生成对抗网络,生成与真实攻击格式一致且内容差异的有效样本。通过采集实车数据作为真实样本进行训练,从特征图、t-SNE可视化、统计学分析和分类器验证角度验证了所生成的增强数据集的实用性,可提高入侵检测分类器准确率;与传统过采样算法含随机过采样(ROS)、合成少数过采样(SMOTE)、SMOTE-ENN、自适应合成过采样(ADASYN)相比,具有更高的准确率。

关键词: 汽车控制器局域网络, 入侵检测, 生成对抗网络, 数据增强

Abstract:

For the data imbalance problem of the current automotive CAN network intrusion detection algorithm due to the lack of attack samples, a CAN intrusion detection data enhancement method based on BEGAN is proposed, which introduces in one-hot coding to image the CAN message features and combines with the constructed Generative Adversarial Network to generate valid samples with the same format as the real attack and with different content. The practicality of the generated enhanced dataset is verified from the perspectives of feature maps, t-SNE visualization, statistical analysis and classifier validation by collecting real vehicle data as real samples for training, which can improve the intrusion detection classifier accuracy. With higher accuracy compared with the traditional oversampling algorithms including Random Oversampling (ROS), Synthetic Minority Oversampling Technique (SMOTE), SMOTE combined with Edited Nearest Neighbors (SMOTE-ENN) and Adaptive Synthetic Oversampling (ADASYN).

Key words: local area network for automotive controller, intrusion detection, generative adversarial networks, data enhancement