| Along with the rapid development of Internet technology, the goals and methods ofnetwork attacks have been richer．The cyber security problems become quite serious. Inorder to improve the security performance of the network and information system,establishing a multi-level and dynamic security defense system is necessary. Intrusiondetection system that can provide a dynamic and real-time protection for the security ofinformation systems turns to be a new field that attracting many researchers’ attentions.However, the huge data that is produced by high-speed network make the intrusiondetection system weak. The complex network environment, the web data becomescomplicated. The result of the increasing of features is the curse of the dimensionalityand over fitting. In order to decrease the computational overhead and improve itsdetecting efficiency of the intrusion detection system, this paper studies the datasetsreducing technology, including features selection and instances selection.The datasets reducing model is constructed on the premise that analyzing theprevious features selection, instances selection and the theory of particle swarmoptimization. Discrete particle swarm algorithm experts at solving complex practicalengineering problems. The datasets reducing model is based on discrete particle swarmoptimization and the encapsulated classifiers. And the model will be detailed designed.Comparative experiments were designed respectively to test the performance of themodel. The feature selection model uses KDD99data sets as experimental data, andcompared with the genetic algorithm-based feature selection and encapsulated classifierto test. The instances selection model uses part of the dataset provided by the UCIinstrument as instance selection training and testing data. And compared with traditional instances reduce algorithm Ordered/Removal. The experiments result show that thediscrete particle swarm algorithm based feature selection can get a smaller subset offeatures but maintain higher classification accuracy, and reducing the time of processingprogress. And can improving data reduction rate on the premise that maintains higherclassification accuracy. |
