Intrusion Detection System Based On Autoencoder And Convolutional Neural Network

Nowadays,the scale of the network is getting larger and larger,generating massive amounts of network data.Traditional machine learning methods need to perform feature engineering on the data set first,and this work requires researchers to have relevant professional knowledge background.In the face of massive and high-dimensional network data,traditional machine learning methods cannot automatically extract features,which will add a great burden to security staff who handle features.Whether the feature engineering is excellent or not will directly affect the performance of traditional machine learning methods.In addition,intrusion detection also has the problem of data imbalance,resulting in poor detection results for a small number of samples.Therefore,how to efficiently and accurately detect abnormal behavior from massive and high-dimensional network data is a hot issue that researchers are concerned about.Aiming at the above-mentioned problems faced by intrusion detection,this thesis proposes an intrusion detection system based on autoencoder and convolutional neural network.First of all,aiming at the problem of high-dimensional data,this thesis proposes a data dimensionality reduction method for stacked sparse autoencoders with attention mechanism.First,through pre-training of each sparse autoencoder,the weights of the stacked sparse autoencoders are initialized,and then fine-tune the overall to obtain a dimensionality reduction model,achieving the goal of mapping the high-dimensional data to more robust low-dimensional data.Secondly,in response to the problem of data imbalance,this thesis proposes a data-level balancing method,BLS-GMM(Border-line SMOTE-Gaussian Mixed Model).We set the number of resampled samples first,and then perform Border-line SMOTE oversampling and GMM undersampling for the categories that are less than and more than the number of resampled samples respectively,so that the system detection rate can be improved.Thirdly,for the problems of low detection efficiency,numerous parameters and easy overfitting,this thesis proposes an intrusion detection system(ICNN-1D-IDS)based on an improved onedimensional convolutional neural network(ICNN-1D).This thesis adopts the adaptive pooling that will make corresponding changes according to the feature map,uses global average pooling and global maximum pooling,and performs cross-layer aggregation to reduce the number of parameters and shorten the training and testing time.Finally,the method proposed in this thesis is evaluated by experiments.Experimental results show that the method in this thesis has very small reconstruction error through stacked sparse autoencoders,and can extract low-dimensional representations of high-dimensional network traffic data.BLSGMM makes the data set in a balanced state to improve the detection rate for classes with a small number of samples.In addition,the intrusion detection method based on the improved onedimensional convolutional neural network has high detection performance,which can reduce the training time and avoid the occurrence of over-fitting.