As the rapidly increasing connectedness of modern vehicles, more and more information security incidents targeting Intelligent Connected Vehicle (ICV) are emerging. Some potential attackers inject malicious packets by external interfaces, which infiltrating the controller area network (CAN), thus implement illegal intrusion. Deep learning-based in-vehicle intrusion detection systems (IDS) among anomaly detection technologies have received a lot attention owing to their high efficiency and accuracy. So, this paper focuses on studying the complex value neural network (CVNN) to detect arbitration field (CAN ID) for protecting CAN network. We proposed an encoder, which can extract shallow features via the auto-encoder algorithm, and furthermore present a random phase that rotates the complex-valued domain features to hide the real features. Then the proposed processing model extracts valuable features with an attention mechanism. Injecting anomaly data in the real vehicle to build the CAN dataset, the real-time detection shows that constructed IDS present a high resulting accuracy, achieving 98%. In particular, the attack experiment indicates that our model makes the adversary hardly inferring valuable information.

随着现代汽车互联性的迅速提高，针对智能互联汽车（ICV）的信息安全事件越来越多。一些潜在的攻击者通过外部接口注入恶意数据包，这些数据包渗透到控制器局域网（CAN）中，从而实现了非法入侵。异常检测技术中基于深度学习的车载入侵检测系统（IDS）由于其高效性和准确性而备受关注。因此，本文重点研究复数值神经网络（CVNN）来检测仲裁字段（CAN ID）以保护CAN网络。我们提出了一种编码器，该编码器可以通过自动编码器算法提取浅层特征，并且还提出了一种随机相位，该随机相位旋转复数值域特征以隐藏真实特征。然后，提出的处理模型通过注意力机制提取出有价值的特征。通过将异常数据注入到实际车辆中以构建CAN数据集，实时检测表明，所构建的IDS具有很高的结果准确性，可达到98％。特别是，攻击实验表明，我们的模型几乎无法使对手推断出有价值的信息。