Intrusion Detection Of Stacked Denoising Autoencoder Deep Neural Networks

With the development of the intrusion behavior,the traditional shallow-seated perception models can't detect complex intrusion behavior.This paper apply deep learning to intrusion detection.Deep learning consists of multi layers formalized neural networks.The weight of deep neural networks is lay-wise initialized by unsupervised learning,then tuned by unsupervised learning.The deep learning does' t only extract more excellent features,but also learn more complex objective funciton.Therefore,deep learning can be apply to process the data that is innumerable and is high dimensional.The advantage of deep learning technology is obvious in intrusion detection,but the training speed of deep neural network is slow.This paper proposes intrusion detection of stacked denoising autoencoder deep neural networks method that prevent overfitting and eliminate noise.This paper mainly include the following work:(1)This paper used the NSL-KDD dataset,but the dataset existed some qualitative features that is discrete digit.This paper adopt high dimensional map method to process these feature in data preprocessing.The data is more distingishable.(2)Because the traditional nonlinear activation function does' t only make the speed of gradient descent slow,but also occur gradient disappearance.This proposed that the hidden used Rectified Linear Units.The experiments showed that the convergent rate of deep neural networks used Rectified Linear Units is faster than simgoid activation function.(3)Because the dataset can be polluted or destructed,this paper proposed that the gaussian noise was added during the learning process of autoencodes to increase autoencoder's robustness.The experiments showed that the accuracy increased about 0.5 percent by the method.(4)The labels of the dataset are not balanced,which easily overfit deep neural network.this paper adopt "Dropout" regularization,which prevent overfitting by similar model combination to prevent outfitting..The experiments showed that dropout was more generic than weight decay.This paper computer the optimal parameters of stacked denoising autoenconders deep neural network by experiments.accuracy rate of the proposal model isvery high in all model that alreadly exists by comparing with other model in the Intrusion detection area.The intrusion of stacked denoising autoencoder deep neural networks has four following advantages: appling to the high dimensional data,increasing generic ablity of deep neural networks,speed up the convergent rate of deep neural networks,increasing the detective accuracy and effectively avoiding misinformation and underreporting.