Research And Implementation Of A Deep-Learning-Based Lightweight Malicious Traffic Identification And Its Distributed Method

With the development of the Internet and communication technology,the network is playing an increasingly important role in social development and people's lives.However,along with the opportunities of development and the convenience of life come challenges that network attacks threatening the Internet all the time.In this situation,an important measure to maintain network security is to monitor and identify malicious traffic.Currently,there are three main problems in the research and application of malicious traffic identification: Firstly,manual intervention preventing the development of rule-based and machinelearning-based identification.How to use deep learning to promote the performance of identificate model is realistic demand of current research.Secondly,although the deep learning can improve the accuracy of recognition,its complex network structure and massive network traffic training datasets makes training to be more difficult.Finally,deep learning model's larger model size and higher model complexity determine the higher demand for the computing and storage resources,which are precise for edge computing equipment.The contradiction gets the application in a fix.In response to the problems above,we propose three optimized and improvement methods:Firstly,we desigin TS-IDS(Temporal-and-spatial-based Intrusion Detection System)model to achieve the goal of better identification performance.Secondly,we implement distributed deep learning technology to achieve the goal of fast model training.Thirdly,we reduce the complexity of the model using pruning technology to achieve the goal of smaller model size.The main contributions of this paper are as follows:(1)We design and implement a lightweight malicious traffic identification system based on distributed deep learning,providing training,optimization,and identification.(2)We propose a new deep learning malicious traffic recognition model TS-IDS,which has higher recognition accuracy and optimized training process of unbalanced data sets.(3)We implement distributed training to speed up the training speed of the TS-IDS,and optimize and adjust the learning rate in a distributed environment.(4)Within network pruning,we reduce the size of the TS-IDS to facilitate the deployment of the model on edge computing devices.