Study On Computational Intelligence Classification Approach And Its Application In Intrusion Detection

Classification, as the term implies in itself, is the process or method of makingirregular things regular. With the rapid development of information technology and theenormous increase of its amount in the form of index numbers, classification has becomean important task in such fields as information processing, data mining and knowledgediscovering. However, the traditional classification methods have demonstrated manyweaknesses and at the same time, intelligence classification method, especiallycomputational intelligence classification method, has drawn special attention and beenwidely used. In this sense, studies in this field have acquired both theoretical significanceand practical value.Intrusion detection, a detection of intrusive action which mainly distinguishes betweennormal network activities and abnormal intrusive ones, is a multi-classification problem inreality. Given that the adoption of computational intelligence classification method canundoubtedly strengthen the effect of intrusion detection, computational intelligenceclassification methods, as well as their application in intrusion detection, are mainlydiscussed and studied in this thesis. Related work and innovations are as follows:(1) To solve such problems as premature convergence and an easy falling into localminima in PSO, CPSO based on cloud model is proposed. The dynamic examination ofinertia weight by adopting cloud model can get a faster optimization speed, avoiding aneasy falling into local minima. After a classical optimization function test, it is concludedthat CPSO is superior to PSO and ACO. Then a neural network classification method basedon CPSO, which can overcome limitations of low accuracy in the neural network algorithm,is also put forward in the thesis. A simulated experiment has shown that this method hasbeen greatly improved concerning classification accuracy.(2) SVM based on statistical learning theories possesses unique advantages inclassification. To solve the limit of experience-based penalty parameter C and kernelparameter in SVM, a cloud model, which can optimize the relevant strategies and increaseconvergence rate, is adopted. The CPSO-SVM, the optimization of SVM model and itsparameter by adopting cloud PSO, is studied. Experiments suggest that in intrusion detection, the accuracy of CPSO-SVM is higher than the classical SVM and PSO-SVM.(3) RVM based on sparse Bayesian framework possesses such advantages as lesscalculation and higher classification accuracy and on the other hand such a disadvantage ofunoptimized parameter. As a result, CPSO-RVM, the optimization of RVM model and itskernel function parameters by adopting cloud PSO, is also studied. Through classicalexperiments and actual detection of multi-classification problems in intrusion detectiondatum of KDDCup99database, it is suggested that compared with such classificationmethods as PSO-RVM, PSO-SVM and CPSO-SVM, CPSO-RVM, CPSO-RVM has thehighest detection accuracy and greatly reduced the rate of misinforming, thus offering ascientific basis for the extensive application of CPSO-RVM.