The Characteristic Analysis And Study On Prediction Technology Of Network Traffic

With the rapid development of computer network technology, the network scale expands rapidly, the bandwidth grows in double, the complexity and heterogeneity increase continually. Applications based on the data network become increasingly widespread, especially the emergence of new applications (such as VoD, VoIP,P2P,etc) leads to a great demand for broadband network services (such as multi-media, video, etc.). As the network traffic unexpectedly sharp increase, the traffic characteristics based on the traditional model is no longer suitable for the current network traffic analysis. Analysis based on the characteristics of network traffic and forecasting technology have great research and application value for guaranteeing the necessary business-critical Quality of Service (QoS), traffic load balancing, differentiated service priority, network detection and location, security attacks (such as DDoS attack), long-term network programming and design.This paper analyzes the network traffic characteristics, including the definition of self-similarity, characteristics of self-similar traffic, Hurst parameter estimation methods which describes the degree of self-similar traffic, basic concept of long-range dependence and heavy-tailed histribution, and the impact of self-similarity, long-range dependence, heavy-tailed histribution to network traffic are described. In the network traffic forecasting, it was discovered that the traditional Poisson model is no longer suitable for network data traffic analysis, a number of new models and traffic forecasting methods have been proposed, at the same time, lots of relevant technology are used to predict traffic, such as neural networks, chaos theory, wavelet methods. We studies previous research on the traffic prediction methods and propose a new combination of network traffic prediction methods based on a variety of forecasting techniques. The method based on theory of decomposition and reconstruction of multi-scale of wavelet, the network traffic, which are decomposed an approximation signal and some detail signals of different scales. Then these signals are reconstructed into several low-frequency and high-frequency time serials by wavelet. These serials are predicted by linear minimum mean square error of estimate (LMMSE) and autoregressive model (AR) respectively according to their different features. The predicted results of all serials are combined into the final prediction traffic. The simulation results with the real traffic traces show that, our method can predict future traffic correctly, and compared with the methods that models for the original traffic and predict directly, our method has a higher forecasting accuracy.Finally, network traffic prediction technology will be applied to distributed denial-of-service attacks (DDoS) detection, this paper presents a denial of service attacks (DDoS) detection model based on the traffic prediction. This model predicts the traffic of business in the future time based on the network traffic forecasting methods referred in this paper, provides a template similar to the normal traffic according to forecast results, to detect whether the network is attacked by DDoS. Because DDoS attack can lead to network traffic burst, and then inflects the self-similarity. As denial of service(DDoS) attacks could lead to a large-scale network traffic mutation, it can be detected rapidly by comparing Hurst coefficient of network traffic under DDoS attacks and the forecasting network traffic.