The challenge of controller area network (CAN) bus topology mapping in black-box environments is addressed in this paper through the proposal of 3Mapping, an unsupervised mapping model based on multi-modal information fusion. The proposed JointVEM multi-modal feature engineering method is used to collect voltage and electromagnetic signals synchronously and to extract a multi-dimensional feature set comprising voltage derivative peaks, electromagnetic peaks, and their temporal intervals from both rising and falling edges. Engine control unit (ECU) hardware characteristics and positional information are effectively integrated, significantly enhancing feature discriminability and robustness. An anomaly data cleaning mechanism based on isolation forest (IF) is introduced to efficiently remove noisy samples. Feature learning in a latent space is implemented using a variational autoencoder (VAE), and unsupervised clustering is performed via a Gaussian mixture model (GMM). Experimental results demonstrate that the JointVEM multi-modal feature method effectively captures node hardware characteristics, achieving silhouette scores above 0.95 and Davies-Bouldin (DB) indices below 0.2. CAN node clustering and bus topology mapping are accurately completed, providing reliable and authentic labels for subsequent security detection.
近年来,很多学者提出针对CAN总线拓扑测绘的研究,和雷达辐射源个体识别类似[10],通过提取节点硬件指纹来区分不同节点。如Schell等[11]提出VALID方法,通过对每条报文采样1~2个显性位电平,并用轻量级递归最小二乘法累积和算法对发动机控制器(Engine Control Unit, ECU)电压特征画像,但该方法的区分能力有限。文献[12]基于时间间隔特征,提出CANvas测绘方法,但对于发送模式相似或由同一定时器触发的不同ECU,其区分能力有限,且容易受到攻击者模仿通信模式的欺骗。
由于不同节点时间间隔或电压特征可能重叠,基于信号传输时延差的测绘方法进入研究者的视野,这类基于信号到达时间差(Time Difference of Arrival, TDoA)的方法,能够直接区分出不同节点的物理位置,提供了更底层、更稳定的区分特征。此类方法的典型代表是TIDAL-CAN(Timing based Intrusion Detection and Localization for Controller Area Network)[14],通过测量同一报文信号到达总线末端两个监控点的差分传播时间来识别不同发送节点。这种方法有效消除了发送时间抖动带来的影响,能够以较高的准确率区分不同位置的节点。其优势在于对替换和插入攻击具有抵抗力。
针对上述问题,提出一种基于多模态融合特征的无监督CAN总线测绘框架(Unsupervised CAN Bus Mapping via Multi-Modal Fusion, 3Mapping),对未知的黑盒CAN总线拓扑进行测绘。实验结果表明,基于多模态融合特征的测绘模型能有效识别不同ECU发送的CAN消息,对消息实现准确聚类,并找到最优聚类数,实现“逻辑节点—消息ID”的有效映射,完成对网络拓扑的测绘。测绘结果能作为监督信号,为样本数据提供真实标签,有效支撑后续检测模型的训练与预测。
首先对 d1数据集进行GMM无监督聚类分析,该数据集包含59 197个样本,对应5个ECU节点发送的13个不同ID的消息。测试结果如图5所示。使用簇内平方和(Within-Cluster Sum of Squares,WCSS)评估聚类效果,并使用肘部法则确定聚类数为5,使用轮廓系数和DB指数也得到同样的结果。进一步检查不同类对应的ID列表,发现不同ECU发送的ID消息被正确聚类为同一类,与CAN总线原型设置完全一致。
SODERIS, COLELLIR, TURRINF, et al. SENECAN: secure KEy DistributioN OvEr CAN through watermarking and jamming[J]. IEEE Transactions on Dependable and Secure Computing, 2023,20(3):2274-2288.
[3]
CUIJ, CHENY N, ZHONGH, et al. Lightweight encryption and authentication for controller area network of autonomous vehicles[J]. IEEE Transactions on Vehicular Technology, 2023,72(11):14756-14770.
[4]
ZHANGG Q, SHENJ, LIJ T, et al. Payload processor: message authentication for in-vehicle CAN bus using data compression and tag filling[J]. Computer Networks, 2025,259:No.111061.
[5]
UMARM, WANGJ D, LIF, et al. Physical layer authentication in the Internet of vehicles through multiple vehicle-based physical attributes prediction[J]. Ad Hoc Networks, 2024,152:No.103303.
[6]
SUNH, SUH B, WENGJ, et al. MTDCAP: moving target defense-based CAN authentication protocol[J]. IEEE Transactions on Intelligent Transportation Systems, 2024,25(9):12800-12817.
[7]
BELLAG, BIONDIP, COSTANTINOG, et al. TOUCAN: a protocol to secure controller area network[C]//Proceedings of the ACM Workshop on Automotive Cybersecurity. New York, USA: ACM, 2019:3-8.
[8]
HAYWARDJ, TOMLINSONA, BRYANSJ. Adding cyberattacks to an industry-leading CAN simulator[C]//Proceedings of the 2019 IEEE 19th International Conference on Software Quality, Reliability and Security Companion. Piscataway USA: IEEE, 2019:9-16.
SCHELLO, KNEIBM. VALID: voltage-based lightweight intrusion detection for the controller area network[C]//Proceedings of the 2020 IEEE 19th International Conference on Trust, Security and Privacy in Computing and Communications. Piscataway, USA: IEEE, 2020:225-232.
[12]
KULANDAIVELS, GOYALT, AGRAWALA K, et al. CANvas: fast and inexpensive automotive network mapping[C]//Proceedings of the 28th USENIX Conference on Security Symposium. Berkeley, USA:USENIX Association, 2019:389-405.
[13]
POPAL, GROZAB, JICHICIC, et al. ECUPrint: physical fingerprinting electronic control units on CAN buses inside cars and SAE J1939 compliant vehicles[J]. IEEE Transactions on Information Forensics and Security, 2022,17:1185-1200.
[14]
MURVAYP S, GROZAB. TIDAL-CAN: differential timing based intrusion detection and localization for controller area network[J]. IEEE Access, 2020,8:68895-68912.
[15]
GROZAB, MURVAYP S, POPAL, et al. CAN-SQUARE-decimeter level localization of electronic control units on CAN buses[M]//BERTINO E, SHULMAN H, WAIDNER M,et al. Computer Security-ESORICS 2021. Cham, Switzerland: Springer, 2021:668-690.
Philips. TJA1050 high speed CAN transceiver datasheet [EB/OL].(2003-10-22)[2025-10-17].
[20]
INSTRUMENTST. SN65HVD23x 3.3-V CAN bus transceivers datasheet (Rev.O)[EB/OL].(2018-04-01)[2025-10-17].
[21]
DENGZ Y, LIUJ J, XUNY J, et al. IdentifierIDS: a practical voltage-based intrusion detection system for real in-vehicle networks[J]. IEEE Transactions on Information Forensics and Security, 2023,19:661-676.
[22]
ZHOUJ, JOSHIP, ZENGH B, et al. BTMonitor: bit-time-based intrusion detection and attacker identification in controller area network[J]. ACM Transactions on Embedded Computing Systems, 2019,18(6):1-23.