Research On Statistical Features For Image Steganalysis Based On Spatial Rich Model

Steganography and steganalysis are contradictory items in the field of information security.The former focuses on hiding secret information in cover objects,and the latter aims to analyze whether there is any embedded secret information in an object.Steganography and steganalysis are of great significance especially for national security,military intelligence,government secrets,the fight against terrorism and confidential business information etc.Image content adaptive steganography hides secret information by slightly modifying the cover(natural)image in rough regions.Image carrying secret information is called a stego image.Qualitative steganalysis is a kind of binary classification problem,and the objective is to distinguish between cover image and stego image.Since the embedding changes may violate the dependencies among neighboring pixels of residual(high frequency component of an image)as well as that of natural image,a popular idea is statistical steganalysis which extracts statistics from residuals to capture the local-range dependency changes caused by embedding.Then,machine learning is employed to train the steganalysis classifier.SRM(Spatial Rich Model)is a typical qualitative steganalysis method.It uses diverse spatial high-pass filtering method to get a variety of residuals;then,the co-occurrence matrix of each residual is calculated and each element of which is the probability of a neighboring residual sample pattern appearing in the residual;finally,the elements of each co-occurrence matrix are rearranged into a vector as the steganalysis feature.In this dissertation,based on the research foundation of SRM,novel steganalysis statistical feature extraction and processing methods are proposed.The main achievements and innovations are as follows.(1)A neighboring residual sample selection method is proposed to improve the performance of SRM feature.The effectiveness of the statistics collected from different types of neighboring residual samples is investigated,and ineffective,effective and high-effective neighboring residual samples are defined.The ineffective neighboring residual samples are not likely to change during embedding,and if they are counted in statistics,they may mix the features with noise.The proposed neighboring residual sample selection method can weaken the interference of ineffective neighboring residual samples.Furthermore,a multi-order statistical feature is proposed to increase the statistical diversity.Experimental results demonstrate that the proposed method can achieve better detection accuracy than SRM.(2)A feature dimension reduction method based on random projection(FPSRM:Fast Projections of Spatial Rich Model feature)is proposed.Increasing the truncation threshold and the co-occurrence order will lead to a higher-dimensional SRM feature which can exploit more statistical bins and capture dependencies across larger-range neighborhood,but this will suffer from the curse of dimensionality.It is noted that each dimension of SRM feature is the probability of a neighboring residual sample pattern appearing in the residual,and each neighboring residual sample pattern can be considered as the coordinates of a point in the high-dimensional space,and all the neighboring residual sample patterns can be viewed as a lattice in the high-dimensional space.The proposed FPSRM method performs multiple random projections on the lattice,and each projection adds up the features corresponding to the neighboring residual sample patterns contained in the same random hyper plane and its adjacent regions to obtain one-dimensional feature.The proposed FPSRM method can capture the transfer of a large number of neighboring residual samples from different directions with different scales and decrease the feature dimensionality.A 7,247,295 dimensional SRM feature and a 658,845 dimensional SRM feature are extracted.Then,those higher-dimensional features are projected into lower-dimensional FPSRM features.The detection performance and the computational complexity of the proposed FPSRM method are investigated.Experiment results demonstrate that the proposed FPSRM feature can achieve better detection accuracy than the typical SRM and PSRM(Projections of Spatial Rich Model),while the computational time cost of FPSRM is much lower than that of PSRM.(3)A feature normalization method(NRS:feature Normalization on Random feature-Subset)is proposed to improve the steganalysis performance of existing steganalysis features while the additional computational time cost is very small.It is noted that the difference between features of cover and the corresponding stego image is very small,while the features of different cover images are often very different.We randomly draw feature-subsets from the existing steganalysis feature,and then these feature-subsets are normalized by using per-sample rescaling method to make them keep the same 1-norm.Two experiments are carried on SRM features.Experimental results show that:under the same dimensionality,the normalized SRM feature can achieve better steganalysis performance than SRM feature;a small amount of normalized feature-subset supplement can rapidly improve the steganalysis performance of SRM feature;compared with SRM feature extraction,the computational complexity of NRS can be negligible.The proposed NRS method can also improve the steganalysis performance of PSRM feature.