| With the rapid development of the network,network security becomes more and more important.Intrusion detection is an important type of network security technology.After decades of development,the existing intrusion detection technology still has deficiencies terms of accuracy and false alarm rate.Neural network-based deep learning methods have powerful modeling and generalization capabilities,which can make up for the above deficiencies.A set of intrusion detection schemes based on the network traffic characteristics dataset KDD99 is implemented in this thesis.Data optimization processing discusses and analyzes numerical coding and onehot coding,Principal Component Analysis and Random Forest processing methods and conducts simulation experiments.An internal nested unit is added to the Long Short Term Memory network(LSTM)to enhance the modeling ability under long time step conditions due to long time dependence of the dataset.To learn the association between adjacent features,1D convolution was conducted to replace the matrix multiplication at the input of the LSTM.These two improvements consist of the C-NLSTM network.The one-hot encoding method and Random Forest feature selection are determined as the data coding and descending dimention algorithms through comparative experiments.With training and predicting the optimized data,the 4hyperparameters of the C-NLSTM network are determined.The accuracy rate of 98.47% and the false alarm rate of 0.7% are achieved in the case of the binary classification,and the detection accuracy of Probe,U2 R and R2 L reaches 99.33%,95.18% and 87% respectively in the case of five classifications.The hardware logic is conducted based on the improved C-NLSTM,including convolution module,dot product module and activation function module and other key calculation modules.Through FPGA function verification,the accuracy rate reaches 91.04% and the false alarm rate is 4.65%,which meets the design requirements. |
PDF Full Text Request