With the network security issues being more prominent, the safety of system andnetwork resources becomes more and more important. Intrude detecting (ID) has become atop research topic nowadays. Support Vector Machine (SVM) has many advantages, such asstronger generalization ability, higher classification accuracy in the classification of thelimited sample of high-dimensional input space. Considering the data and feature redundancyof ID date set, testing will lead to low efficiency and precision. In this thesis, the networkintrude detecting algorithm based on improved SVM has been put forward and it has beentested by KDDCUP1999date set. The main contents are mentioned as below:1) Recursive support vector machine(R-SVM) selects from the results generated bySVM a feature which makes classification performances best to achieve the goal of reducingdimensions. Considering the feature of ID date set, the network ID algorithm based onR-SVM has been proposed. The result shows that SVM sorting algorithm performances betterthan other algorithms mentioned in this article (BP, C4.5, K-neighbor and nearest neighboralgorithms);compare with the classification model under global features, descendingdimension algorithm lowers the training and prediction time;compare with rough setabstracting, R-SVM ensures the same classification quality as the classification under globalfeatures data.2) The basic idea of the local manifold learning is to solve the ‘dimensional catastrophe’through filming the nearby spots on a manifold into instant dots on low dimensionalspace.Furthermore, the adoption of increment learning enables intrude detecting systemenhance its precision along with the accumulation of sample sets.The K-means clusteringmethod obtains a general impression of the raw data, according to such impression, it givesout more rapid and accurate feature extraction and clustering. Aim at the high time-consuming, high and unstable ID data set, a new network ID algorithm based on manifold learning andSVM is applied, along with K-means clustering and SVM increment learning. The resultshows that, compared with the circumstances under no clustering, the K-means not merelyensures the precision but lowers the training and prediction time effectively; compared withR-SVM dimension reducing, manifold learning accelerates the detecting rate and lowers theprediction time; compared with non-increment learning, SVM increment learning properlyimproves the detecting rate and misdeclaration. |