Reliable JPEG steganalysis based on multi-directional correlations

Abstract

With the wide application of JPEG images, JPEG steganography attracts more and more researchers, and accordingly, the detection of JPEG steganography becomes also important. There exist some blind JPEG steganalysis methods, while most of them are either unreliable or time-consuming. This paper presents a reliable and efficient steganalysis scheme to detect the popular JPEG steganography algorithms. First, a novel kind of transition probability matrix is constructed to describe correlations of the quantized DCT coefficients in multi-directions. Then, by merging two different calibrations, a 96-dimensional feature vector is extracted. Additionally, the SVM (Support Vector Machine) is trained to build the steganalyzer. Finally, the proposed feature is evaluated, and a series of experiments are performed on 4 kinds of typical steganography in different embedding ratios, showing that for these steganography algorithms, the new method is more reliable than the best effective blind detection methods existed.