Research On JPEG Image Blind Steganalysis

This dissertation first introduced the history of information hiding, and introduced the most important content of information hiding– the investigation status of image steganalysis, and then on the basis of that, the research of blind steganalysis of JPEG image was in-depth studying. The main content was as follows:(1) A blind steganalysis algorithm using co-occurrence matrix into wavelet domain to reflect the correlation between the coefficients in wavelet domain was proposed. This algorithm first decomposed the image into 3 levels using wavelet transform, and then chose the coefficients from high sub bands to co-occurrence statistic, and after that, Laplace the co-occurrence matrix to pop out the detail of the matrix, and extracted the variance and characteristic function moment of the transformed matrix as feature at last. The experiment results showed that, this method could detect the JPEG stegography (F5, Jsteg, Outguess, Jphide) well.(2) Using the reference of co-occurrence matrix, blind steganalysis algorithm about using joint probability density matrix to reflect the correlation between wavelet subbands was proposed. After decomposing the image into 2 levels using wavelet transform, this dissertation first found the difference of the high frequency sub bands from diagonal, vertical and horizon, and then statistic the adjacency difference matrix using joint probability density, statistic the original image using co-occurrence, and extracted the entropy and energy from the statistic matrix and combined the feature of PDF as the whole features. The result showed that, the detect rate of this algorithm was kept and it had certain generalization ability of MB and PMK.(3) An algorithm based on wavelet inter-subbands and intr-subbands were proposed. First of all, investigate the extracted feature from inter-coefficient in wavelet domain, using difference matrix from vertical, horizon and diagonal into every wavelet subbands and then using co-occurrence matrix to statistic every different matrix, at last extracted entropy and energy as features of coefficients in wavelet subbands. And then, combine the features from inter-subbands and intr-subbands together, and formed 126D feature altogether. The results showed that this feature can detect the tipical steg algorithm (F5, Jsteg, Jphide, Outguess) well, and it had certain generalization ability.At last, summed the whole article, and prospected the next research task...