earticle

영어

Neural Network (NN) models employed in lots of researches about system call intrusion detection, however, most probably suffer from problems, such as sensitivity to random initialization, local optimum, outlier of training data, etc. A robust supervised relevance neural gas with discriminant analysis, called RSRNG-DA, is firstly proposed. By incorporating several robust strategies, e.g. outlier resistant scheme and building discriminant analysis on dissimilarity spaces of prototypes, into supervised relevance neural gas (SRNG) framework, RSRNG-DA possesses better robust properties. RSRNG-DA can tolerate the influence cased by outlier and random ordering of training data, and present a significant improvement on classification efficiency. Moreover, the relevant degree of each system call as feature that contributes most to classification performance can be determined, such that the relevance of all features can help to prune irrelevant pattern dimensions. Our technique is evaluated on the system call database maintained by NMU. Experimental results are compared with other existing methods in the literature, and have shown the superior performance on detection rate, false positive and computation time aspects.