| The richer the application of intelligence and connected vehicles,the stronger the demand for vehicle networking and the more frequent the interaction between the vehicle and the outside world.However,the more interactions,the more serious the exposure of automotive network security vulnerability,and the number of automotive network security attacks have gradually increased in recent years.More and more frequent network interactions lead to a wider network attack surface for intelligence and connected vehicles,and the fundamental of the automobile network is facing increasing penetration of cybersecurity threats.Focusing on the cyber environment of the new generation car network,this paper analyzes the security vulnerability of the network of intelligence and connected vehicles.This paper shows that network security in the car is weak through analyzing the actual cyberattack case and the automotive electronic system architecture.This paper clarifies the vehicle network attack surface and the in-vehicle network attack chain,then statistically classifies vehicle-related network attack troubles in recent years,and shows the corresponding specifications and standards issued by industry and standards organizations.The increasingly serious security threats of automobiles network,this paper focuses on in-vehicle network security and studies in-vehicle network intrusion detection technology based on deep learning.Combining the characteristics of the Generative Adversarial Network(GAN)model,this paper proposes a variety of in-vehicle network intrusion detection technologies based on the GAN model.The main of our work are as follows:This paper is the first to propose single-GAN deep learning models using for in-vehicle network intrusion detection by optimizing the discriminator and generator,respectively.The single-GAN model for in-vehicle network intrusion detection based on the optimized discriminator has optimized neural network,promoted loss feedback,and restructured the loss function.Normal data set and Gaussian noise are used to train the discriminator and generator respectively,and a small amount of attack data participate in the training of the discriminator with a small frequent period for supplemented training,and comparative experiments have reasonably and scientifically conducted.Compared with other GAN-based models in the field of in-vehicle network intrusion detection,the single-GAN model proposed in this paper constructed with less neural network layers is of five classification capabilities,which has a better performance.In addition,the above-mentioned single-GAN model takes 0.12 ± 0.03 milliseconds to detect 64 CAN message frames,and the test accuracy is up to 99.88%.On the basis of the above optimization,this paper proposes a single GAN model for in-vehicle network intrusion detection based on the optimized generator,of which the test accuracy rate can reach 97.96%.In the above model,the loss function of the generator is further modified,and the training participation degree of the attack dataset is gradually reduced,that is,the attack data is only involved in calculating loss feedbacks.Based on the difficulty of obtaining attack datasets and the key role of datasets for training deep learning models,this paper also proposes a step-by-step training GAN-based hybrid model for a situation that the absence of attack data.In the above-mentioned situation,the test accuracy rate of the model is up to 86.61%.During the training process,a small amount of attack data is used to train a redesigned GAN model,and getting a trained generator for training a CNN model,which alleviates the dilemma of lack of attack data to a certain extent.This model provides a new idea for how to train the deep learning model-based for in-vehicle network intrusion detection in the situation of lack of attack data even the absence of attack data. |