Controller area network (CAN) bus-based connected and even self-driving vehicles suffer severe cybersecurity challenges because connections from outside the vehicle and an existing security vulnerability on CAN expose passengers to privacy and security threats. Generative adversarial nets (GAN)-based intrusion detection systems (IDSs) for in-vehicle network can eliminate the limit of insufficient types of attack data suffered by the deep learning-based IDSs. The existing GAN-based IDS is a hybrid deep learning model built by DNN and GAN, which is too complex to have a short detection time. The evaluation performance of this model can be further improved. To mitigate this issue, we propose another GAN-based intrusion detection method for in-vehicle network, which is a single improved GAN. The proposed model can have better evaluation metrics, e.g., the testing accuracy rate is up to 99.8% and poor detection performance is addressed when a single GAN is used in intrusion detection for the in-vehicle network. In this paper, we design a new loss function for generator in GAN to enhance an ability to produce fake abnormal data, and utilize a sparse enhancement training method helping discriminator in GAN to correct the arbitration bias for fake attack data every 100 steps. In addition, we utilize fewer convolution and de-convolution layers for constructing GAN model, which can reduce the calculation time theoretically and ultimately shorten the detection time to [Formula: see text][Formula: see text]ms for a data block built by 64 CAN messages. We evaluate this improved GAN-based intrusion detection by test set. The results demonstrate that our method has not only a capacity of five classifications, but also better evaluation performance than the existing method in the area of GAN-based IDSs for the in-vehicle network.

中文翻译：

---

车联网中使用单一 GAN 的车载网络入侵检测

基于控制器局域网 (CAN) 总线的联网车辆甚至自动驾驶车辆都面临着严峻的网络安全挑战，因为来自车辆外部的连接以及 CAN 上现有的安全漏洞使乘客面临隐私和安全威胁。基于生成对抗网络 (GAN) 的车载网络入侵检测系统 (IDS) 可以消除基于深度学习的 IDS 所遭受的攻击数据类型不足的限制。现有的基于GAN的IDS是DNN和GAN构建的混合深度学习模型，过于复杂，检测时间短。该模型的评价性能可以进一步提高。为了缓解这个问题，我们提出了另一种基于 GAN 的车载网络入侵检测方法，它是一种改进的 GAN。所提出的模型可以有更好的评估指标，例如，测试准确率高达99.8%，解决了单个GAN用于车载网络入侵检测时检测性能差的问题。在本文中，我们为 GAN 中的生成器设计了一种新的损失函数，以增强生成虚假异常数据的能力，并利用稀疏增强训练方法帮助 GAN 中的判别器每 100 步纠正虚假攻击数据的仲裁偏差。另外，我们使用更少的卷积和反卷积层来构建 GAN 模型，理论上可以减少计算时间，最终将检测时间缩短到 [公式：见文本][公式：见文本]ms 构建的数据块64 条 CAN 消息。我们通过测试集评估这种改进的基于 GAN 的入侵检测。