Research On Peer To Peer Trafficidentification Based On Convolutional Neural Netword

Traffic applications based on peer-to-peer(P2P)technology occupy a large amount of network bandwidth and exacerbate the burden on the network.For P2P traffic identification,the traditional model relies heavily on the feature of manual extraction,making it difficult to limit these high-bandwidth P2P applications.More importantly,there is a feature overlap between P2P botnet traffic and normal P2P traffic,which degrades the accuracy of single-phase pattern recognition.This thesis uses deep learning technology to research P2P traffic identification.An improved model that can accurately identify normal P2P traffic and P2P botnet traffic is designed and implemented.The main research results are as follows:1.A P2P traffic identification model based on convolutional neural network is designed and implemented.The model processes load data of a variety of P2P traffic applications and adopts the method of iteratively extracting complex features.Compared with machine learning support vector machine based models,the model has 1%to 3%improvement in accuracy,which reduces the error caused by manual participation.2.An improved solution that can accurately identify P2P botnet traffic is designed and implemented.Aiming at the overlapping features among multiple traffic flows,this thesis designs and implements a two-phase P2P botnet traffic identification scheme.In this thesis,P2P traffic is identified by constructing three kinds of models with coarse granularity,and then P2P botnet traffic is identified by feature extraction combined with classification algorithm.The final recognition accuracy of the whole model is about 97%.3.A model that can accurately identify Distributed Denial of Service(DDoS)attacks in P2P botnets is designed and implemented.The model can imitate the DDoS attack by addding a long-term memory network layer to the convolutional neural network,and accurately identifies low-rate DDoS traffic by using a combination of multiple sets of eigenvectors.Compared with the traditional way of analyzing signaling and features,the model can combine and recognize multiple traffic data,and the final recognition accuracy is about 97.37%.