Multifractal characterization for classification of network traffic.

摘要

This thesis investigates the use of multifractal analysis to characterize network traffic and to facilitate reliable real-time traffic classification. In 1993, a seminal study by Leland et al. revealed the existence a self-affine structure within network traffic. However, despite this discovery, many researchers continued to use traditional techniques of traffic analysis and modelling that did not exploit this knowledge of self-affinity. To demonstrate the general versatility of multifractal techniques to characterize self-affine traffic, this thesis investigates the characterization and classification of a traffic recording from Pear's self-affine data sets which contains an unknown number of classes. To characterize the traffic, the variance fractal dimension trajectory (VFDT) is calculated using a carefully selected window size and window offset. The statistical mean, variance, skewness, and kurtosis are calculated for the VFDT, forming four new statistical trajectories. The histograms of these statistical trajectories are calculated for another appropriate window size, and their stationarity is modelled using the gamma distribution. The resulting eight parameters (two for each of the four gamma distributions) are further reduced to only four parameters using principal component analysis, and the K-means clustering algorithm and Kohonen's self-organizing feature map are used to cluster the data. A locally optimal spread parameter sigma is determined for each probabilistic neural network (PNN) configuration, and a plot of PNN percentage classification accuracy as the number of classes increases reveals that there are most likely three classes in the traffic recording. Finally, an optimized PNN is trained with 50% of the multifractal signatures sampled at regular intervals from the trajectory, and achieves a representative correct classification accuracy of 95% when classifying previously unobserved self-affine traffic.