Research And Applications On Multi Source Data Attack Detection Based On Neural Network

Intrusion detection system(IDS)plays an import role in modern network security system.The core of intrusion detection system is how to detect numerous attack in complex internet environment.Traditional attack detects methods based on feature matching and statistical analysis always facing the problems of feature selection difficulties,poor adaptability and other defects.This lead to low detection rate and high false alarm rate.Neural Network provides a new way for network attack detection.It takes advantages of neural network’s excellent nonlinear expression ability to classify network attacks,so as to overcome shortcomings of traditional methods.1.This thesis presents a comprehensive introduction to attack detection,neural network and deep neural network,and verifies the performance of various neural network models on data sets based on NSL-KDD data set and compares the performance of common full connected neural networks,convolution neural networks and recurrent neural network in data sets.2.A multi-source data attack detection model based on long memory model(LSTM)and deep neural network(DNN)is proposed.The model uses long and short memory model to learn the correlation between data characteristics and data.Deep neural network is used to learn a lot of features of multi-source data.3.The residual neural network(ResNet)is applied to the construction of deep neural network.In the construction of deep neural networks,there are problems of gradient disappearance,over fitting,and difficulty in selecting number of layers.This thesis uses ResNet to construct deep networks.In the structure,batch regularization(Batch Normalization)and Relu are used to mitigate the gradient disappearance and overfitting problems.At the same time,the residual structure can choose a better optimum in order to overcome the problem of layer selection to a certain extent.Based on the NSL-KDD data set,the proposed model is verified,and the performance and time efficiency of the methods such as shallow neural network and deep belief network(DBN)on the test set are compared.The experimental results show that the proposed method of multi-source data attack detection based on LSTM and DNN is 0.856(multi-classification)and 0.914(binary classification)a in recognition accuracy,which is equivalent to the best result of DBN,but the time required for training is only 1/10.Compared with other shallow neural networks,the accuracy rate increased by 0.03~0.05,but the training time did not increase significantly.