Detection, Support Vector Machine-based Network Attacks

Along with the development of business networks, network security has become moreprominent, in particular, attacks on computer systems becomes more complex anddiverse.Network attack detection system is a supplement to traditional computer security,and it increases the system and network security's scope of protection. SVM is a machinelearning algorithm based on the statistical learning theory in the beginning of the 1990s. Itcombines the Maximal Margin Principle and the Kernel Function Theory.It has a solidtheoretical foundation, the ability to promote good and strong nonlinear processingcapacity .And can overcome the "Course Dimensionality", it was paid more attention toand get success in application of pattern recognition and regression estimation.Because of these significant advantages, using SVM method for detecting networkattacks, we can be assured of a good detection capability. This paper has made someattempt on kernel function and kemel parameter selection, includes:1. Against multi-dimensional network, the characteristics of heterogeneous data, webrought the HVDM distance function.Taking into account all the attributes for theclassification of the different contributions to the problem, we used the previousexperience of data, through fuzzy membership function to determine from the variousattributes of different weights and improved SVM RBF kernel function. Thus morereasonable scientific resolved the heterogeneous data discrepancies.2. When a large number of samples involved in the training, the time to find theoptimal parameters grid search algorithm will be consumed long time. Given thesecircumstances, this paper presents a new RBF kernel parameter optimization method,combining feature selection and improved grid search method together to search theoptimal parameters. The results show the proposed method is feasible and effective.