The steganalysis method based on spacing statistics of short duplicate code is an efficient universal detection algorithm for LSB matching steganography.But the steganalysis method must select the appropriate dimension of short duplicate code to meet the different applications.This one-dimensionstatistical analysis method could not take into account the links between multivariate statistical features

thus may affect the detection capability.In this paper

a detection capability law of a single short duplicate code statistical feature is proved

and a method to reasonable choice of the short duplicate code dimension is presented to reduce the detection number.By analyzing the correlation between the statistical features of short duplicate code spacing statistics

a selected feature subset is described as a vector of local features.Then

a universal steganalysis method based on local features extracted from LSB sequences is proposed.The proposed steganalysis method uses the Gaussian Mixture Model (GMM) to describe the multi-dimensional local features

and designs classifier by integrating GMM generative model and SVM discriminative method based on global sequence vocabulary.The experimental results show that under the premise of effective control of the false alarm rate

the proposed method achieves the best overall detection performance to LSB matching steganography and to LSB replacement steganography.