一种融合时序与深度特征的二阶段CAN总线攻击识别方法

谈名名; 张恒; 王鑫; 李明; 张键; 杨明

doi:10.11999/JEIT250651

一种融合时序与深度特征的二阶段CAN总线攻击识别方法

doi: 10.11999/JEIT250651 cstr: 32379.14.JEIT250651

谈名名¹,
张恒^1, ,,
王鑫²,
李明¹,
张键¹,
杨明²

1.
江苏海洋大学计算机工程学院连云港 222005
2.
齐鲁工业大学计算机科学与技术学部济南 250061

基金项目: 国家自然科学基金(61873106, 62303109, 12401679)，江苏省杰出青年科学基金项目(BK20200049)

详细信息

作者简介:
谈名名：男，硕士生，研究方向为信息物理系统安全、网络入侵检测

张恒：男，教授，研究方向为信息物理系统安全、网络安全

王鑫：男，特聘研究员，研究方向为大小模型协同、分布式人工智能

李明：男，讲师，研究方向为大数据分析

张键：男，教授，研究方向为数据挖掘与大数据分析、机器学习

杨明：男，研究员，研究方向为数据安全、物联网安全

通讯作者:
张恒　zhangheng@jou.edu.cn

中图分类号: TN919
计量
- 文章访问数: 36
- HTML全文浏览量: 9
- PDF下载量: 12
- 被引次数: 0
出版历程
- 收稿日期: 2025-07-09
- 修回日期: 2025-09-04
- 网络出版日期: 2025-09-09

A Two-Stage Framework for CAN Bus Attack Detection by Fusing Temporal and Deep Features

1.
School of Computer Engineering, Jiangsu Ocean University, Lianyungang 222005, China
2.
Department of Computer Science and Technology, Qilu University of Technology, Jinan 250061, China

Funds: The National Natural Science Foundation of China (61873106, 62303109, 12401679), The Nature Science Foundation of Jiangsu Province for Distinguished Young Scholars (BK20200049)

摘要

摘要: 控制器局域网(CAN)因安全机制缺失易遭受网络攻击，现有入侵检测系统在多类攻击识别和车载部署上仍存在挑战。该文提出一种融合时序与深度特征的二阶段CAN总线攻击识别方法，通过“先检测、后分类”的策略，将复杂任务分解，实现效率与精度统一。第1阶段设计了数据负载熵(PDE)与ID频率均值偏差(IFMD)特征，从内容与行为2个维度量化报文异常，并利用双向长短期记忆网络(BiLSTM)捕捉时序依赖，实现高效异常检测；第2阶段针对异常样本，引入一维轻量化ParC1D-Net，通过深度特征精细分析实现多类攻击分类。公开数据集实验表明，该方法在Car-Hacking数据集上准确率和F1分数均达99.99%，优于多种先进方法；消融实验验证PDE与IFMD特征在提升异常检测敏感性和鲁棒性方面的关键作用。此外，方法在GPU和模拟嵌入式CPU环境下测试，模型大小仅0.39 MB，实时检测时延分别为0.62 ms和0.93 ms，具备良好部署与实时处理能力。
- 入侵检测 /
- 控制器局域网络(CAN) /
- 二阶段框架 /
- 特征提取
Abstract: Objective The Controller Area Network (CAN), the de facto standard for in-vehicle communication, is inherently vulnerable to cyberattacks. Existing Intrusion Detection Systems (IDSs) face a fundamental trade-off: achieving fine-grained classification of diverse attack types often requires computationally intensive models that exceed the resource limitations of on-board Electronic Control Units (ECUs). To address this problem, this study proposes a two-stage attack detection framework for the CAN bus that fuses temporal and deep features. The framework is designed to achieve both high classification accuracy and computational efficiency, thereby reconciling the tension between detection performance and practical deployability. Methods The proposed framework adopts a “detect-then-classify” strategy and incorporates two key innovations. (1) Stage 1: Temporal Feature-Aware Anomaly Detection. Two custom features are designed to quantify anomalies: Payload Data Entropy (PDE), which measures content randomness, and ID Frequency Mean Deviation (IFMD), which captures behavioral deviations. These features are processed by a Bidirectional Long Short-Term Memory (BiLSTM) network that exploits contextual temporal information to achieve high-recall anomaly detection. (2) Stage 2: Deep Feature-Based Fine-Grained Classification. Triggered only for samples flagged as anomalous, this stage employs a lightweight one-dimensional ParC1D-Net. The core ParC1D Block (Fig. 4) integrates depthwise separable one-dimensional convolution, Squeeze-and-Excitation (SE) attention, and a Feed-Forward Network (FFN), enabling efficient feature extraction with minimal parameters. Stage 1 is optimized using BCEWithLogitsLoss, whereas Stage 2 is trained with Cross-Entropy Loss. Results and Discussions The efficacy of the proposed framework is evaluated on public datasets. (1) State-of-the-art performance. On the Car-Hacking dataset (Table 5), an accuracy and F1-score of 99.99% are achieved, exceeding advanced baselines. On the more challenging Challenge dataset (Table 6), superior accuracy (99.90%) and a competitive F1-score (99.70% are also obtained. (2) Feature contribution analysis. Ablation studies (Tables 7 and 8) confirm the critical role of the proposed features. Removal of the IFMD feature results in the largest performance reduction, highlighting the importance of behavioral modeling. A synergistic effect is observed when PDE and IFMD are applied together. (3) Spatiotemporal efficiency. The complete model remains lightweight at only 0.39 MB. Latency tests (Table 9) demonstrate real-time capability, with average detection times of 0.62 ms on a GPU and 0.93 ms on a simulated CPU (batch size = 1). A system-level analysis (Section 3.5.4) further shows that the two-stage framework is approximately 1.65 times more efficient than a single-stage model in a realistic sparse-attack scenario. Conclusions This study establishes the two-stage framework as an effective and practical solution for CAN bus intrusion detection. By decoupling detection from classification, the framework resolves the trade-off between accuracy and on-board deployability. Its strong performance, combined with a minimal computational footprint, indicates its potential for securing real-world vehicular systems. Future research could extend the framework and explore hardware-specific optimizations.
- Intrusion detection /
- Controller Area Network (CAN) /
- Two-stage framework /
- Feature extraction

HTML全文

图 1 ParC1D-Net模型整体框架图

下载: 全尺寸图片幻灯片

图 2 第1阶段BiLSTM网络模型结构

下载: 全尺寸图片幻灯片

图 3 ParC1D-Net结构

下载: 全尺寸图片幻灯片

图 4 ParC1D Block结构

下载: 全尺寸图片幻灯片

图 5 IFMD 特征窗口大小对模型性能的影响

下载: 全尺寸图片幻灯片

表 1 二阶段模型关键参数设置

模型参数	第1阶段: BiLSTM模型	第2阶段:ParC1D-Net模型
批量大小设置	1024	1024
滑动窗口大小	30	不适用 (N/A)
损失函数	BCEWithLogitsLoss	CrossEntropyLoss
优化器	Adam	AdamW
优化器学习率	0.001	0.001
Epoch设置	5	5
Dropout率	0.5	0.5

下载: 导出CSV

表 2 Car-Hacking数据分布情况

攻击类型	总报文数量	正常报文数量	攻击报文数量
DoS	3 665 771	3 078 250	587 521
Fuzzy	3 838 860	3 347 013	491 847
Gear	4 443 142	3 845 890	597 252
RPM	4 621 702	3 966 805	654 897

下载: 导出CSV

表 3 Challenge数据分布情况

报文类型	总报文数量
DoS	587 521
Fuzzy	196 979
Gear	597 252
RPM	654 897
Attack free	11 742 542

下载: 导出CSV

表 4 Car-Hacking数据集上不同算法间多分类性能比较(%)

算法	年份	准确率	精确率	召回率	F1
Multi-Stage^[10]	2021	99.11	99.13	98.42	99.09
2-Stage Homogeneous^[11]	2022	99.48	99.06	99.21	99.21
HybridSecNet^[12]	2024	99.58	99.58	99.58	99.58
DMLP^[13]	2025	99.95	-	-	99.95
本文方法	-	99.90	99.86	99.54	99.70

下载: 导出CSV

表 5 Challenge数据集上不同算法间多分类性能比较(%)

算法	年份	准确率	精确率	召回率	F1
文献[14]模型	2023	86.08	85.67	86.09	85.88
XGBoost^[15]	2023	98.35	97.54	98.35	97.90
文献[16]模型	2025	95.10	99.60	99.70	96.00
DMLP	2025	99.46	-	-	99.45
本文方法	-	99.90	99.86	99.54	99.70

下载: 导出CSV

表 6 Car-Hacking数据集消融实验结果表(%)

模型	准确率	精确率	召回率	F1
本文模型	99.9 988 ± 0.0 000	99.9 986 ± 0.0 001	99.9 965 ± 0.0 002	99.9 976 ± 0.0 001
w/o PDE	99.9 977 ± 0.0 002	99.9 974 ± 0.0 001	99.9 936 ± 0.0 007	99.9 955 ± 0.0 004
w/o IFMD	99.9 239 ± 0.0 014	99.7 876 ± 0.0 058	99.9 109 ± 0.0 103	99.8 491 ± 0.0 028
w/o PDE&IFMD	99.9 092 ± 0.0 096	99.7 505 ± 0.0 273	99.8 899 ± 0.0 129	99.8 200 ± 0.0 190

下载: 导出CSV

表 7 Challenge数据集消融实验结果表(%)

模型	准确率	精确率	召回率	F1
本文模型	99.9 037 ± 0.0 029	99.8 622±0.0 068	99.5 414±0.0 131	99.7 012±0.0 089
w/o PDE	99.8 743 ± 0.0 060	99.8 201 ± 0.0 077	99.4 017 ± 0.0 323	99.6 098 ± 0.0 185
w/o IFMD	99.8 578 ± 0.0 105	99.8 168 ± 0.0 130	99.3 026 ± 0.0 619	99.5 580 ± 0.0 329
w/o PDE&IFMD	99.8 279 ± 0.0 110	99.7 336 ± 0.0 507	99.2 003 ± 0.0 540	99.4 651 ± 0.0 342

下载: 导出CSV

表 8 不同算法间时空性能比较

算法	年份	模型复杂度(MB)	平均检测延迟(ms)	测试平台
Multi-Stage	2021	3.57	0.02	Intel i5 CPU
TCAN-IDS ^[17]	2022	-	0.05	NVIDIA Jetson AGX Xavier
XGBoost ^[18]	2024	-	33.59	Raspberry Pi 4
DMLP	2025	-	0.17	Raspberry Pi 3B
本文方法(GPU)	-	0.39	0.62±0.10	NVIDIA RTX 3090
本文方法(CPU模拟)	-	0.39	0.93±0.18	Intel Xeon (单核模拟)

下载: 导出CSV

表 9 不同算法间时空性能比较

模型	模型大小 (MB)	推理时延 (ms/样本, bs=1)
单阶段-BiLSTM (多分类)	0.3 128	0.9 300
二阶段-S1-BiLSTM	0.3 118	0.5 628
二阶段-S2-ParC1D-Net	0.0 777	0.6 635

下载: 导出CSV

参考文献(18)

[1]	钱志鸿, 田春生, 郭银景, 等. 智能网联交通系统的关键技术与发展[J]. 电子与信息学报, 2020, 42(1): 2–19. doi: 10.11999/JEIT190787. QIAN Zhihong, TIAN Chunsheng, GUO Yinjing, et al. The key technology and development of intelligent and connected transportation system[J]. Journal of Electronics & Information Technology, 2020, 42(1): 2–19. doi: 10.11999/JEIT190787.
[2]	LAMPE B and MENG W Z. Intrusion detection in the automotive domain: A comprehensive review[J]. IEEE Communications Surveys & Tutorials, 2023, 25(4): 2356–2426. doi: 10.1109/COMST.2023.3309864.
[3]	WANG L X, ZHAO Q C, LEE W B, et al. Deploying intrusion detection on in-vehicle networks: Challenges and opportunities[J]. IEEE Network, 2025, 39(1): 306–312. doi: 10.1109/MNET.2024.3486220.
[4]	WU W F, LI R F, XIE G Q, et al. A survey of intrusion detection for in-vehicle networks[J]. IEEE Transactions on Intelligent Transportation Systems, 2020, 21(3): 919–933. doi: 10.1109/TITS.2019.2908074.
[5]	KHEZRI E, HASSANZADEH H, YAHYA R O, et al. Security challenges in internet of vehicles (IoV) for ITS: A survey[J]. Tsinghua Science and Technology, 2025, 30(4): 1700–1723. doi: 10.26599/TST.2024.9010083.
[6]	LAMPE B and MENG W Z. A survey of deep learning-based intrusion detection in automotive applications[J]. Expert Systems with Applications, 2023, 221: 119771. doi: 10.1016/j.eswa.2023.119771.
[7]	ZHANG H K, HU W Z, and WANG X Y. ParC-Net: Position aware circular convolution with merits from ConvNets and transformer[C]. Proceedings of the 17th European Conference on Computer Vision, Tel Aviv, Israel, 2022: 613–630. doi: 10.1007/978-3-031-19809-0_35.
[8]	SEO E, SONG H M, and KIM H K. GIDS: GAN based intrusion detection system for in-vehicle network[C]. Proceedings of 2018 16th Annual Conference on Privacy, Security and Trust (PST), Belfast, Ireland, 2018: 1–6. doi: 10.1109/PST.2018.8514157.
[9]	KANG H, KWAK B I, LEE Y H, et al. Car hacking and defense competition on in-vehicle network[C]. Proceedings of Workshop on Automotive and Autonomous Vehicle Security (AutoSec) 2021, San Diego, CA, USA, 2021, 2021: 25. doi: 10.14722/autosec.2021.23035. (查阅网上资料,未找到本条文献出版地信息,请确认).
[10]	KHAN I A, MOUSTAFA N, PI D C, et al. An enhanced multi-stage deep learning framework for detecting malicious activities from autonomous vehicles[J]. IEEE Transactions on Intelligent Transportation Systems, 2022, 23(12): 25469–25478. doi: 10.1109/TITS.2021.3105834.
[11]	ALMUTLAQ S, DERHAB A, HASSAN M M, et al. Two-stage intrusion detection system in intelligent transportation systems using rule extraction methods from deep neural networks[J]. IEEE Transactions on Intelligent Transportation Systems, 2023, 24(12): 15687–15701. doi: 10.1109/TITS.2022.3202869.
[12]	CHOUGULE A, KULKARNI I, ALLADI T, et al. HybridSecNet: In-vehicle security on controller area networks through a hybrid two-step LSTM-CNN model[J]. IEEE Transactions on Vehicular Technology, 2024, 73(10): 14580–14591. doi: 10.1109/TVT.2024.3413849.
[13]	ALJABRI W, HAMID M A, and MOSLI R. Enhancing real-time intrusion detection system for in-vehicle networks by employing novel feature engineering techniques and lightweight modeling[J]. Ad Hoc Networks, 2025, 169: 103737. doi: 10.1016/j.adhoc.2024.103737.
[14]	李思涌, 吴书汉, 孙伟. 基于注意力机制的CNN-LSTM网络车内CAN总线入侵检测技术[J]. 信息安全研究, 2023, 9(10): 961–967. doi: 10.12379/j.issn.2096-1057.2023.10.05. LI S Y, WU S H, and SUN W. A CNN-LSTM method based on attention mechanism for in-vehicle CAN bus intrusion detection[J]. Journal of Information Security Research, 2023, 9(10): 961–967. doi: 10.12379/j.issn.2096-1057.2023.10.05.
[15]	JEONG Y, KIM H, LEE S, et al. In-vehicle network intrusion detection system using CAN frame-aware features[J]. IEEE Transactions on Intelligent Transportation Systems, 2024, 25(5): 3843–3853. doi: 10.1109/TITS.2023.3323622.
[16]	陈彦彬, 刘桂雄. 双线性自注意力机制CAN总线入侵检测方法研究[J]. 电子测量技术, 2025, 48(2): 122–130. doi: 10.19651/j.cnki.emt.2417438. CHEN Y B and LIU G X. Study on bilinear self-attention mechanism for CAN bus intrusion detection method[J]. Electronic Measurement Technology, 2025, 48(2): 122–130. doi: 10.19651/j.cnki.emt.2417438.
[17]	CHENG P Z, XU K, LI S M, et al. TCAN-IDS: Intrusion detection system for internet of vehicle using temporal convolutional attention network[J]. Symmetry, 2022, 14(2): 310. doi: 10.3390/sym14020310.
[18]	ALMUTLAQ S, DERHAB A, HASSAN M M, et al. Two-stage intrusion detection system in intelligent transportation systems using rule extraction methods from deep neural networks[J]. IEEE Transactions on Intelligent Transportation Systems, 2023, 24(12): 15687–15701. doi: 10.1109/TITS.2022.3202869. (查阅网上资料,本条文献与第11条文献重复,请确认).