Intrusion Detection Research Based On Improved MobileNet

Network Intrusion Detection（NID）is an important technology for securing networks by dynamically detecting attack-oriented data in network traffic.For unknown or new types of attacks,traditional NID detection may not be effective enough to meet practical needs.On the basis of traditional intrusion detection models,intrusion detection methods incorporating convolutional neural networks（CNNs）can significantly improve the accuracy of classification tasks compared to intrusion detection models incorporating machine learning methods,but CNN models suffer from poor generalisation ability and slow convergence speed during training,leading to low detection accuracy and high false detection rate.In recent years,scholars have increased the complexity of convolutional neural networks to improve the accuracy of the models;however,when complex networks are applied to intrusion detection,their actual detection performance is not necessarily better than that of lightweight convolutional neural network models.To address the problems of poor generalization ability,low accuracy,high false alarm rate and long training time and slow convergence due to high complexity of deep learning models,this paper proposes an intrusion detection method based on improved MobileNet convolutional neural network and completes the following research:（1）Combining a lightweight convolutional neural network model to propose a MobileNetbased intrusion detection The model uses six sets of deeply separable convolutional modules and replaces the fully connected layer with an average pooling layer,which significantly reduces the number of parameters and simplifies the model structure.（2）In order to further improve the classification accuracy and reduce the complexity of the model,this paper proposes an intrusion detection model incorporating MS-IRB and CAM.The model uses three sets of multi-scale inverse residual blocks to improve the efficiency of the model in extracting features from different location data,and introduces channel attention mechanism to enhance the model’s attention to channels containing more valid information,the BN method was chosen to speed up model convergence and reduce overfitting.Finally,the classification results are obtained by averaging the pooling layers and then inputting the results into the Softmax function mapping.Both models proposed in this paper were experimentally evaluated based on the UNSW-NB15 dataset.The results show that the intrusion detection model incorporating MS-IRB and CAM performs better than the traditional machine learning model in terms of accuracy,precision and recall and so on,and the model improves detection accuracy by 17.2% when using 3D data than 2D data.In addition,the model has 34% fewer parameters than CNN,60% fewer parameters than LSTM,and 13% fewer parameters than the MobileNet-based model;and the computation is 45% less than CNN and 26%less than the MobileNet-based model.Therefore,the intrusion detection model proposed in this paper can effectively reduce the complexity of the model and improve the detection effect of the model.