Research Of Intrusion Detection Method Based On RBM-BP Model

With the wide application of the computer and network, the topology of the network is getting more complex and network attack is getting more diverse, which means that we have to find more efficient methods to detect intrusion attack and prevent network abuses. There are two main difficulties during the process of the intrusion detection. Firstly, the traffic data in the network is very complex and unpredictable, accompanied by noise data. This means that the intrusion detection model has to possess the strong ability to prevent the noise data and to learn new features of traffic data accurately. Secondly, the methods and patterns of network attack keep changing. And with rapid growth of the network traffic data, the previous model may not have a nice recognition ability. Hence, a desirable and effective intrusion detection model has good ability to learn the new patterns of network attack by itself and can fast process complex high-dimensional data. In current studies the current intrusion detection techniques mainly include intrusion detection based on statistics, intrusion detection based on data mining, intrusion detection based on machine learning, intrusion detection based on the artificial neural network.As one of the typical machine learning algorithms, artificial neural networks are widely applied in intrusion detection. BP neural network is the most classic neural network model, and it has a strong learning ability and nonlinear ability. But when facing the complex high-dimensional data, BP neural network may learn slowly, and can’t work efficiently, which means it can’t get a good classification accuracy. On the other hand, Restricted Boltzmann Machine(RBM) has a good ability to learn the characteristics and patterns of original data autonomously, and it can reduce the dimension of high-dimensional data.Therefore, this paper proposed an intrusion detection method based on RBM-BP.As for the complex and high-dimensional data, this paper took advantage of RBM with a strong feature learning ability to reduce the data dimension by eliminating redundant features and noise data through unsupervised learning on those data. Then,this paper utilized BP neural network which has a strong nonlinear ability to sever as the classifier to ensure the accuracy of intrusion detection. The effectiveness of the RBM-BP model has been proved via experiments. First, this paper tested the classification performance of the proposed model according to accuracy rate, false positive rate and missing report rate. Then, this paper tested the proposed model’s recognition capability for each type of network data under the condition of multiple classifications. Last, this paper compared the performance of the RBM-BP model and the Fisher-BP model and further verified that the RBM-BP has a better performance.This paper put forward the RBM-BP model by taking the advantage of RBM and BP respectively. Not only can it save time without no manual label and reduce dimensions of data, but also it equips with a considerable high accuracy, a desirable low false positive rate and missing report rate.