A Lightweight IoT Intrusion Detection Method Based On One-dimensional Convolutional Neural Network

The implementation of 5G has promoted the development of the Internet of Things(Io T).With the widespread adoption and growing importance of the Internet of Things,it can easily attract the attention of hackers and become a prominent target.Therefore,network intrusion detection for Internet of Things has been important.Previously,many scholars have proposed intrusion detection algorithms for the Internet of Things based on neural networks,but most of these algorithms have complex network structures and a huge amount of parameters,which are not suitable for deployment to Internet of Things terminals.Therefore,the terminal node needs to upload the data to the detection center,which brings additional communication costs and new attack risks.To this end,the thesis explores how to construct lightweight Io T intrusion detection methods.In order to reduce the feature redundancy of the dataset and the computational pressure of intrusion detection,the thesis firstly proposes an improved genetic algorithm feature selection mechanism based on the Pearson correlation coefficient.Compared with the basic genetic algorithm feature selection mechanism,this method controls the population dispersion degree in the population initialization stage and improves the convergence speed of the algorithm.Finally,under the action of this method,the dimension of the dataset decreased by 90.7% compared with that before feature selection,the accuracy of intrusion detection increased by 13.4%,and the alarm rate increased by 27.8%.Moreover,the training time of the intrusion detection algorithm model is reduced by 33.1% under the same training conditions with unselected dataset.On the basis of feature selection,in order to further realize the lightweight goal of intrusion detection algorithm,the thesis proposes a lightweight Io T intrusion detection algorithm based on one-dimensional convolutional neural network.The algorithm uses a one-dimensional convolutional neural network as the basic model,and then applies the hierarchical weight pruning algorithm for further compression and optimization.Compared with the basic model,when the detection effect is similar,the size of the model is reduced by 52.4%,the number of parameters is reduced by 10 times,and the inference time is reduced by 55.7%.And on the binary classification task on the UNSW-NB15 dataset,the proposed algorithm outperforms most of the recent related works and is comparable to the best among them.Finally,based on the previous two works,the thesis proposes a system for building a lightweight intrusion detection model for the Io T.It's a highly scalable,visualized and easy-to-operate platform for the construction of lightweight neural network intrusion detection models,and it realizes the whole process management of dataset import,feature selection,data preprocessing,model pre-training,model lightweighting and intrusion detection and verification.