Intrusion Detection Algorithm Based On Deep Belief Networks

With the extensive popularization and application of the network, the attention of network security continues to heat up, which has already risen to the national strategic level. Intrusion detection technology is an important part of network security and it is an effective measure of network security. Traditional intrusion detection technology has the problems of a single detection method, poor detection performance and low adaptive capacity. Due to the increasing espansion of the network, the diversification of network attacks and new vulnerabilities, the traditional intrusion detection technology can not meet the needs of network security. Deep belief networks is a cutting-edge technique in the fields of machine learning with the advantages of the nonlinear network structure such as being able to extract essential features having achieved remarkable results in the field of computer vision, speech recognition, natural language processing. It is extremely urgent to introduce deep belief networks into the intrusion detection field to solve the problems of the current network security. The work mainly done in this paper is listed as follows:(1)Proposing an intrusion detection algorithm based on adaptive deep belief networks.Intrusion detection requires a high adaptive ability. Sampling for contrastive divergence algorithm of deep belief networks is easy to fall into local optimal value and a problem of being sensitive to learning rate parameters when in deep belief networks training. In order to solve the above problems, an intrusion detection algorithm based on adaptive deep belief networks is proposed. First, from the perspective of statistical mechanics, the enery change trend of the component Restricted Boltzmann Machine model of deep belief networks is analyzed. Secondly, in order to overcome the fixed rate of experiential learning causing system instability problem, learning rate based on changes in energy optimization strategy is proposed. Finally, by combining parallel algorithm and optimization strategies, an intrusion detection algorithm based on adaptive deep belief networks is proposed. This algorithm can effectively improve the detection rate of intrusion detection and so other indexes.(2)Proposing an intrusion detection algorithm based on deep belief networks and random forest.Due to various types of intrusion detection attacks, difficult feature extraction of high dimensional network data, and high rate of false alarm, along with the advantages of deep belief networks and integration algorithm, an intrusion detection algorithm based on deep belief networks and random forest is proposed. In this algorithm, the feature extraction is carried out by using the multi-layer structure, the high dimensional network data mapped to the low dimension space. In the feature extraction process, characteristics of different hidden layers are extracted. Finally, random forests used combinatorial feature to detect the intrusion. The experimental results show that the proposed method can effectively improve the detection rate and reduce the rate of false alarm.(3)Proposing a hybrid intrusion detection model for unbalanced network data.Aiming at problems of imbalanced distribution of attacks from network data and low detection rate of minority class attacks by traditional intrusion detection algorithm, a novel hybrid intrusion detection model is proposed. In the preprocessing model of model data, the improved Synthetic Minority Over-sampling Technique(Improved Trigonometric Function Synthetic Minority Over-sampling Technique, ITFSMOTE) is introduced for minority class data.In the synthesis of minority class data, the method of triangle selection and the thought of survival of the fittest are adopted, which increases the probability of selecting a small number of data. Intrusion detection module in the hybrid intrusion detection model uses the improved intrusion detection algorithm in the paper to identify the type of attacks. Experimental results show that the proposed model can effectively improve the detection rate and reduce false alarm of minority intrusions.