Research On Feature Selection For JPEG Steganalysis

As a novel solution of secret communication, information hiding can conceal secret information and make those unauthorized not suspect the existence of the message. However, once the technology is applied illegally, it will harm the country’s and the pepole’s benefit. Therefore, reserches on steganalysis are of great significance, which can not only detect the existence of information, but also obtain the content of information. Many typical steganalysis algorithms generate a large number of features relative to the size of the training set. High dimensionality of the feature space can reduce classification accuracy, obscure important features for classification, and increase computational complexity.This paper focuses on feature selection based on four steganography algorithms, including F5, Jsteg, MB1and Outguess, of images stored in the JPEG format.This paper firstly realizes MSFR algorithm, which is the feature selection algorithm using the Mahalanobis distance for steganalysis. By feature ranking via Mahalanobis distance, the optimal feature subset is selected. Four steganography algorithms with different message lengths are used in the experiment, and the experimental results show that for JFMP-274algorithm, feature dimensionality can be reduced to3%of the full suite of274features, that is10features. However, the time complexity of MSFR algorithm is high and the time is about21hours.For the high time complexity of MSFR algorithm, this paper presents SMSFR algorithm. Firstly, according to the redundancy of the features, the algorithm obtains sub-optimal feature subset, and then the time complexity can be reduced using MSFR algorithm. The experimental results show that the optimal feature subsets at various levels of embedding of four steganography algorithms can be used to give comparable results to the full suite of274features. Plus, SMSFR algorithm can reduce the time complexity compared to MSFR algorithm and the time is about3hours.In order to reduce the time complexity of MSFR algorithm effectively, this paper presents two algorithms:PCA-MSFR algorithm and LDA-MSFR algorithm, which respectively use the methods:principal component analysis and linear discriminant analysis. PCA-MSFR algorithm firstly uses PCA to obtain the sub-optimal feature, and then uses the MSFR algorithm to reduce the feature dimensionality. The experimental results show that compared to MSFR and SMSFR algorithm, the time complexity of PCA-MSFR algorithm can be reduced. However, the accuracy rate results of PCA-MSFR algorithm are lower than SMSFR algorithm’s. LDA-MSFR algorithm uses LDA to obtain the sub-optimal feature, and then to use the MSFR algorithm to reduce the feature dimensionality. The experimental results show that the optimal feature subsets at various levels of embedding of four steganography algorithms can be used to give comparable results to the full suite of274features. Plus, the time complexity is greatly reduced. And the universal steganalysis method, MSFR-36algorithm, obtained by the LDA-MSFR algorithm, is effective for four steganography algorithms above. At last, for Yass steganography algorithms of images stored in the JPEG format, the experimental results show that for JFMP-710algorithm, feature dimensionality of710can be reduced to40features.