Quantitative Steganalysis Algorithm For Image Steganography

Steganography is a technique of obtaining stego object by embedding secret information into digital cover. It can transfer stego objects through an open channel without any suspicion, so it is widely applied in many fields. It also could be used to transfer military secret or dangerous information by lawbreaker. Steganalysis is the countermeasure technology of steganography, with the purpose of detecting the presence of the secret information, estimating the length of secret information and extracting it from stego image. Steganography can effectively prevent the abuse of steganography, and has attacted a great deal of attention.This paper fouse on the technology of imagesteganalysis, and the main purpose is estimating the secret message length. This paper proposes a universal steganalysis and further proposes three quantitative steganalysis algorithms:1. Steganalysis based on co-occurrence Matrix in DCT Domain. Typical steganography methods maintain the first order statistics unchanged. The co-occurrence matrix describes the transition probability among coefficients, which is the second-order statistical method. It can effectively reflect the changes before and after embedding information. The co-occurrence matrices of low frequency alternative current (AC) coefficients in DCT domain are taken as image features in this paper. With the Support Vector Machine (SVM) as classifer, stego images can be detected effectively.2. Quantitative steganalysis based on image features and support vector regression (SVR). Majority steganalysis algorithms can only detect the existence of hidden message, but the estimation of hidden message length is more important. This paper proposes an improved universal quantitative steganalysis for estimating the hidden message length. Firstly, the 132-dimension features, including the Markov statistics of the inner and among DCT blocks, are extracted in the DCT domain. SVR is used to establish the mapping model between features and embedding rates of stego images. The simulation results reveal that the proposed quantitative steganalysis is a feasible and effective method for estimating the embedding ratio of F5 and outguess stego images in practice.3. Universal quantitative steganalysis based on multiple-domain Features and Partial Least Squares Regression (PLSR). This scheme extracts image features in the spatial domain, DCT domain and DWT domain respectively. PLSR is used to establish quantitative steganalyzer by learning the relationship between features and embedding rates. The simulation results show that the proposed scheme is of fast speed and good performance for predicting the embedding rates of secret message in F5, outguess and MB1 stego images.4. A universal quantitative steganalysis based on wavelet Hidden Markov Tree (HMT) and PLSR. HMT model in the wavelet domain fully consider the clustering inner scales and the persistence among scales, the statistical dependencies and non-Gaussian statistics of wavelet transform coefficients. The wavelet domain HMT model parameters are taken as image features, PLSR is utilized to establish the quantitative steganalyzer. Stego images embedded by F5, Outguess and MB1 are used to testing the quantitative steaganalyzer, and the simulation results demonstrate that the quantitative steaganalyzers can estimate the message embedding rates of accurately and fast.