The prior knowledge for traditional steganalysis, such as steganography algorithms, embedding rates and sources of images, etc., is difficult to be satisfied in practice. In the scenario of blind steganalysis that the above conditions are unknown, analysis using clustering can effectively distinguish between the actor who performs steganography and the others. We propose a method for fusion which is suitable for the selected features, and is to improve the accuracy of JPEG's steganalysis via clustering. It fuses the principal components of the feature based on partially ordered Markov models with the feature based on calibration, and makes full use of complementarity between features as well as reduces the redundancy, identifies out of the guilty actor better and improves the accuracy of identifying actors who perform steganography. Experimental results show that by different steganography approaches and in different embedding rate conditions, using our scheme can obtain a general increase in the accuracy of JPEG steganalysis by about 2% compared to the existing methods, and get a highest accuracy up to 16%.
