<p>Description of Car Hacking Dataset.</p>

The emergence of Autonomous and Connected Autonomous Vehicles (CAVs) has transformed the automotive landscape drastically over the past few years by offering enhanced features in the vehicles for drivers’ safety and convenience. These developments have introduced various features in AVs i.e., lane-keeping, cruise control, etc. These features are mainly powered by the Electronic Control Units (ECUs) that communicate using the Controller Area Network (CAN) bus protocol. The components in the AVs communicate with each other by sending and receiving messages via the CAN bus. However, despite increased connectivity, these vehicles have become vulnerable to cyber attacks, as malicious actors can exploit the CAN protocol to manipulate vehicle behavior, which can not only threaten the safety of the passengers but public as well. Hence, several Intrusion Detection Systems (IDS) have been proposed, however, these systems struggle with computational complexity, limited effectiveness against sophisticated attack types, and a lack of interpretability and transparency of detection mechanisms. To address challenges in the existing systems, this paper presents a novel hybrid Deep Learning (DL)-based IDS using DL components such as Convolutional layer and Long Short-Term Memory (LSTM) layers to capture complex patterns in the CAN messages. The proposed IDS uses a residual connection to enhance gradient flow and training stability. The system is evaluated on four common attack types, namely RPM Spoofing, Gear Spoofing, Fuzzy, and Denial of Service (DoS), achieving a detection accuracy of 99.99%. Finally, the outcomes of the proposed IDS are visually interpreted using the Explainable AI (XAI) technique called Local Interpretable Model-agnostic Explanations (LIME) to provide transparency into the model’s decision-making process, thus increasing trust in the system’s deployment in real-world AV environments.

过去数年间，自主驾驶与联网自主车辆（Autonomous and Connected Autonomous Vehicles, CAVs）的兴起极大重塑了汽车产业格局，为车辆搭载了诸多可提升驾驶员安全与出行便利性的功能。此类技术发展为自主车辆（Autonomous Vehicles, AVs）引入了车道保持、巡航控制等诸多功能，这些功能主要由电子控制单元（Electronic Control Units, ECUs）提供支撑，而各单元通过控制器局域网（Controller Area Network, CAN）总线协议进行通信。自主车辆内的各组件通过CAN总线收发消息实现相互通信。然而，随着连接性不断提升，此类车辆也愈发容易遭受网络攻击：恶意攻击者可利用CAN协议操纵车辆运行状态，这不仅会威胁乘客安全，还会危及公共安全。为此，诸多入侵检测系统（Intrusion Detection Systems, IDS）已被提出，但现有系统仍存在计算复杂度高、对复杂攻击类型检测效果有限，以及检测机制缺乏可解释性与透明度等问题。为解决现有系统的诸多挑战，本文提出了一种基于深度学习（Deep Learning, DL）的新型混合入侵检测系统，该系统采用卷积层（Convolutional layer）与长短期记忆（Long Short-Term Memory, LSTM）网络层等深度学习组件，以捕获CAN消息中的复杂模式。所提出的入侵检测系统引入残差连接以增强梯度流动与训练稳定性。该系统在四种常见攻击类型——即转速欺骗（RPM Spoofing）、挡位欺骗（Gear Spoofing）、模糊攻击（Fuzzy）与拒绝服务（Denial of Service, DoS）——上进行了评估，检测准确率达99.99%。最后，本文采用名为局部可解释模型无关解释（Local Interpretable Model-agnostic Explanations, LIME）的可解释人工智能（Explainable AI, XAI）技术，对所提入侵检测系统的结果进行可视化解读，以阐明模型的决策逻辑，进而提升该系统在真实自主车辆环境中部署的可信度。