Network Intrusion Detection Based on Stacked Sparse Autoencoder and Binary Tree Ensemble Method

Abstract

With the increasing of network attacks, the traditional machine learning method can not solve the classification problem of massive intrusion data effectively. This paper proposes a Xgboost based on stacked sparse autoencoder network(SSAE-XGB) method to learn latent representation of original data. Due to inconsistent category distribution of training and test dataset, we use the sparsity constraint to enhance the generalization ability of the model. Stacked sparse autoencoder network is employed to reduce the dimension of high-dimensional and unlabeled original data, so as to obtain the deep feature representation of the original data. Due to the class imbalance of intrusion data, this paper proposes a novel hybrid classifier, which is constructed by using binary tree and ensemble method. Our experiments with all NSL-KDD dataset demonstrate that our proposed SSAE-XGB binary tree and ensemble method can achieve incredibly high performance in terms of F1 and it outperforms the previous work.