The Universalsteganalysis Based On Statistical Manifold Reduction

With the popularization of computers and Internet, Information Technology isplaying more and more important role in people’s daily life. As a result, it bringsconvenience to people’s work and life, whereas it also causes the emergence ofsecurity risks. That is the reason why people have paid more attention to informationsecurity.Steganography is an important branch of information security. The principle of itis to achieve invisible communications by making use of the redundancy ofmultimedia which can embed secret messages. In order to avoid being used byterrorists, steganalysis, a technique corresponding to steganography, comes toappearance. The goal of steganalysis is to detect whether the multimedia file acting asa carrier contains secret message. As a result it has an adversarial relationship with thesteganography. We mainly talk about universal steganalysis algorithms in this paper,the universal steganalysis aims at detecting a variety of steganorgaphy algorithms.The most simple kind of universal steganalysis can determine whether the multimediadata contains secret messages. It is very necessary to take further research onsteganalysis for the reason that it is the key technique to prevent the illegal use ofsteganography.Generally, the procedure of universal steganalysis algorithm can divide into twosteps: feature vector extraction and classification, since the vector has a tendency tobe high dimensions. With the consideration that it is difficult to deal with highdimensional data which can cause a problem of dimension disaster, as a result, it isnecessary to take dimension reduction on that high dimensional data.We propose a statistical manifold dimension reduction algorithm in this paper,for the purpose reducing high dimensional feature vector and keeping a higherclassification rate in the lower dimensional feature. Statistical manifold dimensionreduction is a kind of nonlinear dimension reduction, every point in this manifold isparameterized by PDF(probability density function). The procedure ofthis algorithm is as follow:1. Using fisher information metric measures the difference among probabilitydensity function of image feature vector, considering the probability density functionas a point in the manifold space； 2. Computing the geodesic distance matrix along the manifold. To get the matrix,we need the result of step1；3.Embedding the geodesic distance matrix to lower Euclidean space by LEMdimension reduction, in another word we can describe the higher data in another waywhich can represent the nature of the data.The experimental results show that this algorithm is effective to manysteganography methods including JSteg, F5, nsF5, MBS1and MBS2. It candistinguish cover image and stego image in a higher accuracy, moreover it is usefulfor a variety of information hiding algorithm. As a result, it is feasible and practicableto use statistical manifold dimension reduction in steganalysis field.