Research On Steganalysis For LSB Matching Based On Image Local Features

As an important subdiscipline of information hiding, steganography is to communicate in a concealed way which hides the existence of the secret information. The steganography by an image has become a very hotspot because of the extensive application of images. Steganalysis, as the opposite technology against steganography, is to detect, extract, restore and destroy the secret messages embedded into the cover images. Doing researches on steganalysis can not only test the security performance of steganographic methods, but also supervise the multimedia data in the network for protecting the safety and stability in our country and society.For the statistical features of images with different contents are dissimilar, the detection of the concealed messages is influenced by the statistical features of secret messages, the embedding mechanism, as well as the features of the given cover images. Compare with the stego images which have been compressed or coded, steganalysis for LSB matching embedding in uncompressed high-resolution images is much more difficult, especially in condition of low embedding rate.This thesis focuses on the study of steganalysis for LSB matching embedding based on the image content. In the thesis, the connection between the local features of images and the statistical changes caused by LSB matching is analyzed, and several LSB matching steganalysis methods are proposed. The contributions obtained in this thesis can be summarized in the following aspects:1. A LSB matching steganalysis method based on the local properties of histogram is proposed. The changes of deviation and difference features of images with different texture complexity before and after LSB matching embedding is analyzed. Based on the conclusion that the changes are larger in smooth areas, 18 features, including deviation and difference local features are extracted, and the Fisher linear discriminator is applied for classifying. Experimental results show that the proposed method exhibits good performance in Camera database.2. The influence on different bit planes by the LSB matching embedding is analyzed. The value of incomplete gamma function is computed according to the approximate entropy parameter to measure the randomicity of different bit planes sequence, which is affected by texture complexity of an image. The analysis indicates that the LSB matching steganography would have different influence on different bit planes, the higher, the little.3. A method for attacking LSB matching steganography is proposed based on the correlation of local pixels in bit planes. The process of LSB matching embed is modeled as a corruption of additive noise to the"pure"cover image, and the image restoration is introduced to reconstruct the cover image by median filter. Considering that images can be assumed to be local stationary sources, the least three significant bit planes are taken as the main object. The alteration rate of the image blocks with the same gray levels defined by Huang is extracted to judge whether there has been embedded any secret messages or not. Experimental results show that the proposed method exhibits good performance in UCID database. 4. A method aimed at detecting LSB matching is proposed based on the difference features of bit planes. The analysis indicates that the least three significant bit planes are influenced much stronger than others after embedding. Consequently take them as a new image. The local features of difference histogram and co-occurrence matrix of the new image are put into fisher linear discriminator. Experimental results show that the proposed method exhibits excellent performances for the detection of LSB matching steganography at high embedding rate in both Camera and UCID database. The detecting rate is improved when the embedding rate is low as well.Finally, the research work is summarized and further research topics and directions in the future are discussed.