Research On IoT Intrusion Detection Method Based On Heterogeneous Transfer Learning

The Internet of Things(IoT)is more widely used in our life because of its interconnection characteristics.However,IoT faces many challenges and problems,and the information security problems of insufficient computing resources and network attacks are lunched.Aiming at the heterogeneity of IoT devices,the lack of computing resources,and the insufficiency and imbalance of data sets,this thesis studies the mapping deviation between the source domain and target domain when the target domain label is unknown based on heterogeneous transfer learning.The research works are shown as follows:(1)Aiming at the problem that IoT devices have limited ability to learn abnormal patterns,which leads to the inability to actively carry out deep learning,an improved transfer learning method of VGG(VGG Reduce,VGG-RED)IoT intrusion detection model is proposed.Firstly,the similar features of five datasets were analyzed,and the public features and private features were extracted and dimensionality reduced through the preprocessing methods of data cleaning and data conversion.Secondly,VGG model was used to independently supervise the learning of private features on five intrusion detection datasets.Then,a transfer model was obtained by training the source domain data and fine-tuning the model structure and parameters.Finally,the weights of intrusion detection features trained by the VGG-RED model were transferred to the IoT intrusion detection application by using the transfer method.Experimental results show that in terms of performance,the model reduces the calculation time by 8%～13%.The results show that the proposed model can not only solve the problem of heterogeneity and imbalance of IoT device data,but also improve the accuracy of detecting malicious attacks in IoT.(2)Aiming at the problem of domain shift faced by domain adaptation,a multisource domain feature unsupervised domain adaptation network is proposed.The problem of domain shift is that when the same type of IoT device monitors different attacked devices,the performance characteristics of abnormal data collected may be different.Moreover,when the network has multiple source domains,the invariant representation of the common domain is difficult to unify,and there are large differences between the feature Spaces of data categories in different domains,which leads to the misclassification of samples near the class boundaries.Therefore,this thesis proposes an IoT intrusion detection method based on transfer learning and multi-source unsupervised domain adaptation,which uses a three-stage alignment strategy.(1)Feature distribution alignment: the feature distribution of source domain and target domain in multiple specific feature Spaces is aligned;(2)Conditional distribution alignment: aligns the conditional distribution of label categories for each pair of source and target domains;(3)Classifier output alignment: domain confrontation is used to confuse the source domain and the target domain,and the domain-specific classifiers are aligned so that all source domain classifiers have the same probability of output for the same target sample.The migration experiments on the IoT and industrial Internet of things intrusion detection classification benchmark datasets N-Ba IoT and Edge-IIoTset show that the average accuracy is improved by about 4%and 27%respectively.This shows that the new method can not only solve the problem of multi-source domain data offset,but also improve the accuracy of detecting IoT attacks and the ability of anomaly classification,and has good generalization.The method proposed in this thesis effectively improves the problems of lack of IoT device resources and insufficient data sets,realizes low-latency and efficient detection of anomaly detection,and designs the corresponding three-stage domain alignment strategy to promote the solution of multi-source domain offset data characteristics.This thesis has 20 figures,12 tables and 105 references.