Research On Intrusion Prevention Based On Quantum-behaved Particle Swarm Optimization And Semi-supervised Clustering

Traditional network security technologies are firewalls and intrusion detectionsystems, but these methods have some disadvantages. As an important method toinsure the computer network security, intrusion prevention technique becomes a hotresearch topic in the information security field.Firstly, an intrusion detection algorithm based on semi-supervised clustering withQuantum-Behaved Particle Swarm Optimization (QPSO) is presented. It can solvesome problems, such as the low detection rate of unsupervised learning algorithms, andthe low detection rate of unknown intrusion behaviors and the insufficiency of trainingsamples of supervised learning algorithms. And to over the defects of traditional FuzzyC-means (FCM) algorithm that can be only used to deal the spherical datasets andsensitive to the noised and wild value data, we propose a new objective function whichincorporates Bray-Curtis distance to deal the spherical and non-spherical shapeddataset or heavy noised dataset, and this method is considering the additional penaltyterm to avoid poor clustering result.Secondly, with development of computer technology, a lot of mass data areemerging. These data lead to many learning algorithm face with the problem of“Dimensionality Curse”. So a new semi-supervised clustering algorithm based onsemi-supervised dimensionality reduction is proposed. The semi-supervised fuzzyclustering algorithm with pairwise constrains recently proposed by Grira is analyzed.Because of the disagreement on the magnitude order between constraint term andobjective function of competitive clustering algorithm(CA) and the competitive term inthe objective function lacks intuitive explanations，aiming at this problem，an improvedalgorithm is proposed based on a redefined objective function. Its penalty cost functionintroduces a new co-expression of two samples in the pairs, which has the samemagnitude order as that of the typical fuzzy clustering; Through adding Renyi entropyinto competitive term, it can prevent premature convergence and accelerate convergenceas one nears the right number of cluster, Furthermore, KDD CUP99data set isimplemented to evaluate the proposed algorithm. Compared to other algorithms, theresults show the outstanding performance of the proposed algorithmAt last, we present an intrusion prevention system model based on QPSO and semi-supervised clustering. It includes detection system module, Intrusion PreventionSystem (IPS) responce module, message register module and central control module.We design the detection algorithms of the intrusion detection module, and use thesemi-supervised clustering along with QPSO as the detection algorithms. Function andconfiguration of the other modules are presented.