Research On Virus Detection Model Based On Artificial Immune System

With the coming of the "Super Internet" and "4G",computer viruse breaks out more frequently.In current study,detection methods for computer virus are in the stage with traditional feature matching,which have many problems to be solved urgently,mainly reflected in the lack of anti-confusion ability,the lack of intelligence,the lack of correlation analysis and incapable of action for the new unknown virus etc.To solve these problems,in this thesis,motivated by the ideas of negative selection algorithm,clonal selection algorithm and hybrid intelligent algorithm in artificial immune system,three kinds of computer virus detection models are designed respectively.Firstly,in the third chapter,we propose a virus detection model based on negative selection theory and graph pattern mining,which uses DFS encoding sub-graph to show action invokes characteristics.In this model,we improve the algorithm of feature extraction,improve the g-Span algorithm,and design a function to calculate two graph's relation to get better efficiency of two sub-graph pattern comparison.The experimental results show that the average detection accuracy(DR)of the model is 97%,and the average detection accuracy(ACC)is 97%.Secondly,in the fourth chapter,we focus on the direction of family association analysis,propose a complete algorithm scheme based on clonal selection,and design a computer virus detection model combined with clustering algorithm.In the detection of malicious viruses,it also gives the characteristics of its family behavior,and provides a large number of reference information for subsequent virus processing.Experiments show that the average value of model identified malicious virus correctly rate(SR)is 94.62%,the family gives the correct reference features included for the average rate of 98.16%,and a number of experimental group appeared in the right contains a ratio of 100%.Finally,in the fifth chapter,we design a virus detection model based on artificial immune neural network,and give a variety of universal optimization strategies and solutions.The experiments show that when the size of training set data is over 100000,the average correct recognition rate(TDR)is 92%,and the false detection ratio(FPR)maintains at about 8%.