Network Intrusion Detection System Based On Multi-scale Convolutions

Traditional network intrusion detection systems that are based on rule bases are unable to detect new attacks,and machine learning algorithms need manually extracted features of network data.Manually extracting features have failed to identify constantly emerging attacks while deep learning algorithms can automatically extract features of given data.This paper proposes a network intrusion detection system based on deep learning algorithms and the main work is as follows:1)Several network intrusion detection data sets are compared and analyzed.The UNSW-NB15 data set is selected as the data to validate the model.After analyzing the data set,the data are processed.Symbolic features are encoded into numerical features,then network flows are converted into feature maps that conform to the input format of convolutional neural networks after data normalization.2)Classification results of the traditional convolutional neural networks Le-Net5 and Alex Net on UNSW-NB15 data set have shown several problems.Aiming to solve the problems of classification accuracy and imbalanced data,solutions are proposed.The modified Gradient Harmonizing Mechanism loss function is applied in the training phase of convolutional neural networks.Also a convolutional neural network with multi-scale convolutions is constructed.LSTM(Long Short-Term Memory),batch normalization and global average pooling are used to optimize the performances of proposed model.The model performances under different network parameters are also analyzed.3)The proposed model is evaluated through being compared with other machine learning and deep learning algorithms.Results have shown that our model can achieve an overall higher accuracy rate and can detect minority samples more accurately.The model has also achieved outstanding results on the CICIDS2017data set,and a simple network intrusion detection system is designed based on the data set.