Research On Network Intrusion Detection Based On GRU_SVM

With the deepening integration of the Internet and society,network security issues have become a hot issue that people are now concerned about.Intrusion detection technology is one of the important guarantees of network security operation.Although a large number of scholars have done research on intrusion detection technology,there are more or less problems in those studies.The detection rate is still low,especially for small sample attacks,so we can't detect effectively attack behavior.In order to solve these problems,an intrusion detection method based on GRU?SVM model is proposed.This method uses the Tensorflow learning framework as a platform to build a training model,and performs preprocessing suchas digitization and normalization of the intrusion detection dataset by adopting Pandas,Numpy and other function libraries.By using the excellent feature extraction and learning capabilities based on the gated recurrent units(GRU)network,combined with the easy fit of the multilayer perceptron(MLP),feature learning is performed on the preprocessed data and its vectors are reduced to low dimensions.In order to improve the detection rate of the detection system,support vector machine(SVM)is used to detect and classify the reduced dimension attacks.In this experiment,two common intrusion detection datasets are selected,the NSL-KDD dataset and the CICDDo S2019 dataset.The experimental results of different training epoches and learning strategies were analyzed,and by adjusting the parameters of GRU?SVM model,such as the number of neurons,batchsize and so on,we can get better performance.Analyzing and comparing the experimental results of GRU,LSTM,GRU?MLP,LSTM?MLP,Bi GRU?MLP,Bi LSTM?MLP,the results show that when using the NSL-KDD dataset,the detection rate of large sample attack behaviors(DOS and PROBE)is basically unchanged,while for the small sample U2 R attack type,the change from scratch is realized,and the small sample R2 L attack type is detected.The overall detection rate is significantly higher than that of a single GRU model.When using the CICDDo S2019 dataset,compared with a single GRU model,the detection rate of various attack types has been significantly improved.