Research On Abnormal Network Traffic Detection Based On Deep Learning

With the rapid development and extensive application of the new Internet,the network scale,network structure and network application are becoming more and more complex,and the network carries more and more traffic.In order to enable the whole industry to enjoy the huge dividends created by the safer Internet,it is necessary to ensure the safe and efficient operation of the network,so can quickly and accurately detect the abnormal network traffic is a powerful means to improve the security of network communication.In recent years,with the development of artificial intelligence,machine learning has been widely studied and applied in more and more technical fields.As a subset of machine learning,deep learning is widely used in speech recognition,automatic machine translation,instant visual translation,automatic driving and other fields with its advantages in efficiently processing complex and high-dimensional data.In the study of highly non-linear network traffic data,traditional learning methods(such as Adaboosting,SVM and random forest)used in previous studies are difficult to improve the performance of identifying network anomalies,and deep learning has not given full play to its advantages in this field.In this paper,four deep learning models constructed based on FCN(Fully convolutional network),VAE(Variational Auto encoder),VAE-FCN(combined application of VAE and FCN)and Seq2Seq(sequence to sequence)are designed.The application of four advanced deep learning models in abnormal network traffic detection is still in its preliminary stage.This article aims to study the feasibility and performance of each model,and does not conduct in-depth research on other aspects such as how to adjust parameters.This paper selects five representative public data sets for experimental analysis.The experimental results confirm the feasibility of applying deep learning to abnormal network traffic detection.In particular,the Seq2Seq detection model has the best performance evaluation results in all data sets.,Its accuracy is as high as 99%,and the NSL-KDD data set that is often used for research is more than 15% higher than traditional machine learning methods.