A Novel Hybrid Quantum-Classical Framework for an In-Vehicle Controller Area Network Intrusion Detection
A Novel Hybrid Quantum-Classical Framework for an In-Vehicle Controller Area Network Intrusion Detection
复制标题
DOI：
10.1109/access.2023.3304331
复制
发表时间：
2023
期刊：
影响因子：
3.9
通讯作者：
M. Salek;P. Biswas;Jacquan Pollard;Jordyn Hales;Zecheng Shen;Vivek Dixit;M. Chowdhury;S. Khan;Yao Wang
中科院分区：
文献类型：
--
作者：
M. Salek;P. Biswas;Jacquan Pollard;Jordyn Hales;Zecheng Shen;Vivek Dixit;M. Chowdhury;S. Khan;Yao Wang
车载控制器局域网 (CAN) 由于其基于广播的通信特性，很容易受到各种网络攻击。攻击者可以通过无线通信、信息娱乐系统或车载诊断端口向车辆的 CAN 注入虚假消息。因此，有效的入侵检测系统对于区分真实的 CAN 消息和虚假消息至关重要。在这项研究中，我们使用经典神经网络 (NN) 和量子受限玻尔兹曼机 (RBM) 开发了混合量子经典 CAN 入侵检测框架。经典神经网络致力于从车辆 CAN 总线数据生成的 CAN 图像中提取特征。相比之下，量子 RBM 专用于 CAN 图像重建，以实现基于分类的入侵检测。该研究的新颖之处在于利用 RBM 的生成能力来重建 CAN 图像中的像素，其中一部分专用于标记。然后，重建图像的该部分用于将图像分类为攻击图像或正常图像。为了评估混合量子经典 CAN 入侵检测框架的性能，我们使用现实世界的 CAN 模糊攻击数据集来创建三个独立的攻击数据集，其中每个数据集代表与车辆相关的一组独特特征。我们将混合框架的性能与类似但仅限经典的框架进行了比较。我们的分析表明，与纯经典框架相比，混合框架在 CAN 入侵检测方面表现更好。对于本研究中考虑的三个数据集，混合框架中的最佳模型分别实现了 97.5%、97% 和 98.3% 的入侵检测准确率以及 94.7%、93.9% 和 97.2% 的召回率。相比之下，仅经典框架中的最佳模型分别实现了 92.5%、95% 和 93.3% 的入侵检测准确率以及 84.2%、89.8% 和 88.9% 的召回率。
In-vehicle controller area network (CAN) is susceptible to various cyberattacks due to its broadcast-based communication nature. An attacker can inject false messages to a vehicle’s CAN via wireless communication, the infotainment system, or the onboard diagnostic port. Thus, an effective intrusion detection system is essential to distinguish authentic CAN messages from false ones. In this study, we developed a hybrid quantum-classical CAN intrusion detection framework using a classical neural network (NN) and a quantum restricted Boltzmann machine (RBM). The classical NN is dedicated to feature extraction from CAN images generated from a vehicle’s CAN bus data. In contrast, the quantum RBM is dedicated to CAN image reconstruction for classification-based intrusion detection. The novelty of the study lies in utilizing the generative ability of an RBM to reconstruct the pixels in a CAN image, a portion of which is dedicated to labeling. Then, that portion of the reconstructed image is used to classify the image as an attack image or a normal image. To evaluate the performance of the hybrid quantum-classical CAN intrusion detection framework, we used a real-world CAN fuzzy attack dataset to create three separate attack datasets, where each dataset represents a unique set of features related to the vehicle. We compared the performance of our hybrid framework to a similar but classical-only framework. Our analyses showed that the hybrid framework performs better in CAN intrusion detection compared to the classical-only framework. For the three datasets considered in this study, the best models in the hybrid framework achieved 97.5%, 97%, and 98.3% intrusion detection accuracies and 94.7%, 93.9%, and 97.2% recalls, respectively. In contrast, the best models in the classical-only framework achieved 92.5%, 95%, and 93.3% intrusion detection accuracies and 84.2%, 89.8%, and 88.9% recalls, respectively.