Multi-classifier Intrusion Detection Based On Deep Belief Network Under Unbalanced Data

Intrusion detection is an active security defense technology that can effectively identify the intrusion behavior in the network.However,the number of different attack types in the existing network varies greatly and the network data exhibits massive and high-dimensional complex features,this leads to the learning-based intrusion detection model is not effective and it is difficult to identify the types of attacks with a small number of samples in the network.In view of the above problems,the use of Gradient Boosting Decision Tree in ensemble learning for good classification performance of unbalanced data,combined with Deep Belief Networks in deep learning for superior processing performance of massive and high-dimensional data.The processing performance,a multi-classifier intrusion detection model based on deep belief network is proposed(DBN-MGBDT).Firstly,the data of the intrusion detection data is standardized and normalized to form a sample dataset.Secondly,the DBN model trained by the sample dataset is used to extract low-dimensional and representative characterization data from massive and high-dimensional intrusion detection data.At the same time,the feature data is binary classified according to the label corresponding to the data.Finally,the one-to-one method is used between any two types of feature data to construct a classifier using GBDT to perform final classification and identification of network data.Using the NSL-KDD dataset,the experimental comparison with the existing DBN-BP,DBN-MSVM,DBN-SOFTMAX,SVM and GBDT models shows that the evaluation indexes of the DBN-MGBDT model are stable and superior;For the randomly selected groups of unbalanced experimental data,the average accuracy and detection rate can be as high as 99.13% and 98.63%,which is better than DBN-BP and DBN-MSVM model about 2%;For a small number of R2 L and U2 R attack types,the average detection rate is 91.62% and 75%,also have better detection performance for a smaller number of attack types.The dissertation contains 20 figures,17 tables and 51 references.