JPEG Steganalysis Based on Classwise Non-Principal Components Analysis and Multi-Directional Markov Model

Abstract

This paper presents a new steganalysis scheme to attack JPEG steganography. The 360 dimensional feature vectors sensitive to data embedding process are derived from multidirectional Markov models in the JPEG coefficients domain. The class-wise non-principal components analysis (CNPCA) is proposed to classify steganograpghy in the high-dimensional feature vector space. The experimental results have demonstrated that the proposed scheme outperforms the existing steganalysis techniques in attacking modern JPEG steganographic schemes -F5, Outguess, MB 1 and MB2.

Universal statistical steganalytic method

“…The next verification of proposed steganalytic method was focused on its comparison with existing image steganalytic methods. Proposed method used an Ensemble classifier and statistical vector with the length SS (285+46) was compared with current image steganalytic methods published in [5] and [7].…”

“…The first compared method [5] is based on the extraction of statistical vector with the length 360, while the main part consists of Markov model statistical parameters from transition matrices. The classification into stego or cover object class was performed by CNPCA classifier [6].…”

“…Steganalysis is utilized to detect a presence of steganography system. The first method [5] is based on the extraction of statistical vector composed by 360 parameters, while the main part consists of Markov model statistical parameters from transition matrices. 120 parameters are extracted from the transition matrix and JPEG images are scanned not only by "zig-zag" form but also vertical and horizontal scanning are utilized.…”

Steganalysis for Markov Feature of Difference Array in DCT Domain

“…Markov matrix was employed to capture the correlations existed in mode 2-D arrays. By utilizing both intra-block and inter-block feature, the steganalysis capability has been enhanced, while compared with individual intra-block features [3,4]. However, the detection performance became poor while attacking F5 with low embedding rate.…”

“…Difference arrays along horizontal and vertical directions in JPEG 2-D array were calculated and Markov matrix is applied to modeling these difference 2-D arrays to construct intra-block feature. In [4], Xuan introduced different local scanning orders and bidirectional Markov matrix to investigate intra-block dependencies. It had demonstrated good classification performance and computational efficiency.…”