Design And Implementation Of Network Anomaly Detection System Based On Deep Learning

In today’s rapidly changing internet technology,the network has become an essential part of people’s work and life.Although internet technology brings great convenience,there are also some security issues that cannot be ignored.It not only affects the interests of individuals and enterprises but also relates to international security.Network anomaly detection is an important research direction in the field of network security.Deep learning methods have shown great potential in the field of network anomaly detection with their powerful feature learning ability and robustness.In this thesis,we have studied and explored intrusion detection and deep learning-related knowledge and proposed an intrusion detection model based on the attention mechanism of Bi GRU-CNN network,and further improved the effect of the model using a cost-sensitive matrix loss function.The main work of this thesis is as follows:(1)The research background and significance of network intrusion detection are introduced,and the research status at home and abroad is understood,laying a good foundation for the follow-up research.(2)In-depth research on existing intrusion detection related knowledge,and detailed introduction of related knowledge.(3)Proposed an intrusion detection system based on Bi GRU and CNN.The model uses Bi GRU to extract data features,fully considering the characteristics of the feature before and after,and using attention mechanism to extract significant features.Using the model in UNSW-NB15 dataset,the multi-classification accuracy is89.89%,which is better than some other experiments —— classical machine learning algorithms such as SVM and deep learning algorithms such as CNN network and LSTM network.(4)Detailed introduction of optimization ideas for the problem of imbalanced experimental data,and using resampling techniques and a cost-sensitive loss function that can be co-trained for training.This model in the UNSW-NB15 dataset has a multi-classification accuracy of 90.03%,and higher accuracy and lower FAR on the classification of multiple minority classes.