Research On Universal Steganalysis Of Image For Low Embedding Rates

With the rapid development of internet, information security technique hasbecome increasingly important. Through more than10years’ development,information hiding technology based on the multimedia content has also beenconcerned by academia and practitioners. The relationship of between steganalysisand steganography in information hiding is not only incompatible but alsocomplementary to each other. This thesis is focused on the techniques of lowembedding rate steganalysis in digital images. Two universal steganalysis algorithmsfor different format images are presented. The main content of this thesis issummarized as follows:By studying steganographic algorithms for JPEG images and existing algorithmsof feature extraction, a new method of low embedding rate universal steganalysis inDCT domain is proposed for dealing with the problem of blind detection for JPEGimages. DCT coefficients of the DCT coefficient blocks and of rearranged DCTcoefficient blocks correlation is utilized to construct different feature matrix.Different feature vectors allocate different weights to enhance the effect of theclassifier. Moreover, several kinds of typical lower embedding rate steganographicimages in DCT domain are detected by using SVM as a classifier. The simulationresults show that when the embedding capacity of above1KB, the correct rate ofstego images judgments can reach above95%.At present the most existing steganalysis based on multi-domain features forBMP images received little attention and have the disadvantages of low detection rate.A new method based on multi-domain features of low embedding rate universalsteganalysis for BMP images is presented. The features are extracted from gradientenergy differences in spatial domain, DCT coefficients correlation in DCT domain,the mean and the standard deviation of difference values matrix in DWT domain. Toreduce the features difference of an image itself and enhance the effect of theclassifier, an image was predicted to format its corresponding prediction image. Thesimulation results show that compared with the existing methods, the detectionachieves better reliability when the embedding capacity of above2KB.