Intrusion Detection Algorithms For Vehicle Networks Based On Improved Autoencoders

With the continuous development of the automobile industry,in-vehicle networks play an increasingly important role in information transmission,status monitoring,and networking anti-theft.However,when the in-vehicle network provides various services to people,it also means that a vehicle increases the risk of being attacked by the network and has more security risks.How to ensure the security of the in-vehicle network has become a significant work.In recent years,due to the excellent performance of deep learning in various fields,researchers have begun to combine deep learning with intrusion detection algorithms for in-vehicle networks.At present,the Deep Learning-based Vehicle Network Intrusion Detection(DL-VNID)algorithm has many shortcomings while having high performance.For example,in the Software Definition Vehicle Network(SDVN)of DL-VNID,due to the insufficient ability of the automatic encoder(Automatic Encoder,AE)to extract the normal traffic characteristics in SDVN,AE has a low classification accuracy in SDVN intrusion detection.In addition,due to the problem of catastrophic forgetting in the incremental learning process of DL-VNID,the classification accuracy of the deep learning model for the old category data is not high.Aiming at the above-mentioned problems,this paper proposes an enhanced intrusion detection algorithm and an incremental intrusion detection algorithm.The main work of this thesis is as follows:(1)An intrusion detection algorithm which uses Strengthen Asymmetric Automatic Encoder based on semi-supervised learning(SAAE)is proposed to improve the classification ability of automatic encoders in SDVN intrusion detection.SAAE consists of two parts: Abnormal Model Warehouse(AMW)and Time-Convolutional Automatic Encoder(T-CAE)based on temporal feature maps.In the training phase,this paper first uses a small amount of abnormal traffic to build AMW,which preserves different kinds of abnormal traffic features.Then AMW and T-CAE are jointly trained using normal traffic,and the neural network parameters of AMW and T-CAE are simultaneously updated through the loss function proposed in this paper during the joint training process.In this paper,two different AMW structures are proposed respectively,and a different loss function is proposed for each AMW.In the testing stage,the accuracy rate of the single T-CAE neural network on the intrusion detection dataset In SDN can reach95.60%,while the classification accuracy of T-CAE after different AMW enhancements on the In SDN dataset can reach 98.32% and 98.99%,respectively.At the same time,by comparing SAAE with existing intrusion detection algorithms,this paper proves that the SAAE algorithm has the advantages of high accuracy and low false alarm rate in SDVN intrusion detection.(2)An incremental intrusion detection algorithm which use Collaborative Incremental Learning based on Confidence(CILC)is proposed to overcome the catastrophic forgetting problem of DL-VNID in incremental learning.The CILC algorithm consists of two parts based on the Asymmetric Multi-feature Fusion Automatic Encoder(AMAE)and Classification Deep Neural Network(C-DNN).First,in the incremental learning phase,the CILC algorithm trains an AMAE model and a C-DNN model for each batch of new data.In the detection stage,when there is only one C-DNN and one AMAE model in CILC,the output of C-DNN is the final classification result.When there is more than one C-DNN,CILC will pass the input data through all AMAE and C-DNN,and then use the result of AMAE as the confidence level to select the output of a certain C-DNN as the final result.In addition,this paper uses Variational Automatic Encoder(VAE)to oversample sparse samples,and prioritizes sparse samples in the detection stage to improve the detection accuracy of CILC algorithm due to unbalanced data sets.In the experimental stage,this paper uses the CICIDS2017 data set and the CAN-intrusion data set as the network environment outside the vehicle and the network environment inside the vehicle to test the effectiveness of the CILC algorithm.Experimental results show that the CILC algorithm can not only retain the ability to classify old categories,but also has a high classification accuracy for new categories of data.At the same time,the CILC algorithm also has a good classification ability for some categories with sparse samples.