Research Of Feature Reduction And Traffic Classification Method Based On SVM

Network traffic feature reduction and classification can realize the differential management according to the application protocol.They are the foundation of network protocol design,network operation management and network traffic scheduling.They can provide a method for network security detection and traffic cleaning.The paper starts with the traffic feature reduction and classification method based on Support Vector Machines(SVM),and mainly includes the following two aspects: One is to solve the performance degradation of classification model,which is caused by high dimensionality or redundancy of network traffic features.The Filter and Wrapper feature selection methods are combined to improve the evaluation criteria and search strategy of feature selection.At the same time,based on the feature extraction and sample space transformation theory,the feature extraction model is researched which embedded in the feature selection module.The other is to solve the equilibrium problem of SVM model's empirical risk and the generalization performance.Based on nonlinear SVM kernel function transformation theory,parameter optimization methods of SVM kernel function are analyzed,classification and convergence ability enhancement methods of optimization algorithms are researched.The main research contents and achievements are as follows:1.For selecting the optimal feature subset which can represent the original traffic data distribution characteristics,a Filter-Wrapper mixed feature selection model is proposed.The model firstly investigates the contribution of a feature to traffic classification based on the Filter mode,and removes the features which are smaller than the set threshold according to the weight of each feature in the original feature set.Then,on the new feature subset,the model adopts Wrapper mode to perform two-time feature selection based on SVM and the corresponding search strategy.Finally,the model selects a feature subset with strong discrimination ability.The model solves the problem that the combination features with strong distinguish ability are deleted by only using Filter mode.And it solves the deviation between the feature evaluation result and the final classification algorithm.2.Aiming at the problem that SVM is susceptible to redundancy features with high similarity dependence,a feature extraction model of Principle Component Analysis with two-time feature selection module is proposed.The model uses two-time feature selection module adaptively when it need identify key features.It checks the correlation of each feature,locks the key feature with the corresponding search strategy,and strengthens the maximum correlation-minimum redundancy of the features.Through the sample space transformation,the model can obtain the feature subspace in a certain direction,and can reduce the data input width,reduce the computational complexity and shorten the training time effectively.The experimental results show that the effect of model's dimension reduction is obvious.On the small sample data set,the model can achieve the classification results consistent with the original sample data set which has practical significance for real-time network traffic classification.3.So as to balance the empirical risk and generalization performance of SVM classification model,the SVM parameter optimization methods are researched on,and two improved parameter optimization algorithms are proposed respectively.One is the improved grid search parameter optimization algorithm(IGS).IGS can dynamically adjust the two-time search area,reduce the grid density,improve the efficiency of search algorithm,and prevent overfitting phenomenon when optimizing parameters.The other is the improved particle swarm optimization algorithm(IPSO).IPSO uses nonlinear inertia weight and asynchronous optimization learning factor to balance global and local search ability.In the process of evolution,IPSO can find the optimal solution quickly.The experimental results show that within the finite computation,the proposed parameter optimization algorithms of the classification model can find the optimal parameter combination from the parameter space.They improve the classification and generalization ability of SVM effectively,and can achieve good classification performance on all six different SVM training models.4.In order to reduce the two-time feature selection process,accurate interpret the key components of original feature combination,to solve the parameter optimization algorithm early maturity and elite individual loss problem,a traffic classification model(GB-SVM)with feature selection and bacterial foraging parameter optimization is built.In the process of feature selection,the model based on genetic algorithm(GA)to select the optimal feature combination,it does not require two-time feature selection,is able to lock the key features of network traffic,and takes into account the correlation between the data.At the same time,the model can interpret the components of original key feature accurately.In the process of parameter optimization,an improved bacterial foraging parameter optimization(IBFO)algorithm is proposed.The algorithm improved the core operator of IBFO,which does not need a wide range of search,does not need to be adjusted,and is not easy to appear early maturity.The improved core operator can ensure the local search ability of the algorithm,avoid the loss of elite individuals at the same time,and converge to the global optimal solution more quickly.The experimental results show that the building model time of GB-SVM is significantly shortened,and the traffic classification accuracy is significantly improved.Compared with other typical supervised classification methods,GB-SVM has obvious advantages in classification performance and generalization ability.