Research On Intrusion Detection Method Based On LightGBM Feature Extraction And Deep Learning

In recent years,the scale of Internet access,business traffic and complexity have all undergone tremendous changes.With the expansion of networks and systems,security threats are becoming more and more severe.Network intrusion detection is one of the hotspots in the field of information security and network security.Deep learning can effectively improve detection performance,but it is limited by the scale,quality of the data set,and the current mainstream deep learning models take a long time to train and poor results.This paper studies the network intrusion detection algorithm based on deep learning.First,the LightGBM algorithm is used to balance the data distribution,and then an intrusion detection algorithm based on the attention mechanism of the two-way gated recurrent unit(Attention-BiGRU)neural network model is proposed.Finally,the algorithm effectiveness of is verified.(1)Compare and analyze the structure and characteristics of typical intrusion detection systems,as well as analyze the ideas and processes of commonly used intrusion detection algorithms The three data sets used in the experiments in this article are preprocessed,including one-hot encoding,processing of non-numerical,processing of missing values and infinite values,and de-dimensionalization processing.(2)In the multi-class detection,the deep learning model does not perform well in the minority class detection of the data set.We start processing with the data set,first use the LightGBM algorithm to extract the features of each category of the data set,and then use BSMOTE(Borderline-SMOTE)for the minority class Over-sampling,NM(Near Missing)under-sampling processing is adopted for most classes to ensure a more balanced data set distribution.(3)Propose an intrusion detection algorithm based on the combination of attention mechanism and BiGRU,design an intrusion detection model,and verify the detection performance of two-class and multi-class respectively.The results show that the detection accuracy rate of the algorithm in the NSL-KDDtest+ data set is 93.8%,the detection accuracy rate in the NSL-KDDtest21 data set is 72%;the detection accuracy rate and the recall rate in the UNSW-NB15 data set are 93.8%,respectively And 58.5%;the detection accuracy and recall rate in the CICIDS2017 data set are 97.4% and 91.4%,respectively,which are higher than several current mainstream detection algorithms.In the multi-class detection,combined with the new data set constructed by the LightGBM algorithm,the data distribution of various attack behaviors is balanced,and it is applied to the training of the Attention-BiGRU neural network model,which aims to improve the detection accuracy of the minority class and reduce the training time.