Quantitative Steganalysis Based on Wavelet Domain HMT and PLSR

Aiming at the problem of estimation of secret message length in steganalysis, this paper presents a quantitative steganalysis method based on HMT (Hidden Markov Tree) and PLSR (Partial Least Squares Regression) to solve the problem. In this paper, three 2-State HMT models are modeled respectively for wavelet coefficients in the horizontal, vertical and diagonal directions. In order to calculate the parameters of HMT, EM (Estimation and Maximization) algorithm is adopted to train the HMT models. The parameters are used as the 66-D feature of image. Then, the quantitative steganalyzer which is used to estimate the message length is established by combining HMT with PLSR. The proposed scheme is evaluated by constructing quantitative steganalyzers for F5, outguess and MB, simulation results demonstrate that these quantitative steaganalyzers can estimate the message embedding rates accurately and fast.