The Improved RSM Algorithm For High-dimensional Image Steganography Detection

Image steganography blind detection is a pattern recognition process, including feature extraction and classification construct two aspects. Due to the high dimensional steganography detection feature can help capture more information about steganography, therefore, in order to combat increasing complexity and sophistication of modern image steganography techniques, blind steganography detection algorithm inevitably need to use more and more high-dimensional eigenvectors. In particular, in order to confront the newly proposed HUGO steganography, steganalysis extract feature vectors up to tens of thousands. The use of high-dimensional characteristics of the classifiers steganography structure is a challenge:it will not only increase the complexity of the classifier training, and even affect the ability of the classifier. Currently, in order to solve the problems caused by high dimensional feature, scholars mainly use feature dimension reduction techniques. Common feature dimension reduction techniques are often required to operate global features, while global features on the high dimensional feature set is an expensive operation work, therefore, the general dimension reduction technique is not applicable to confront HUGO. Random Subspace Method (RSM) is a classical ensemble learning algorithm, which is from the original high-dimensional space extract low-dimensional feature space subset, then in each low-dimensional feature subset to build the base classifiers. Training at low dimensional space can greatly reduce a single base classifier training complexity. Random subspace method is particularly applicable to train classifier in high-dimensional sample space, therefore, this article will approach the improved methods on the basis of the classical random subspace, the main works include:(1) The improved RSM based on double Bagging sampling and chi-square statistic. In order to achieve good classification performance of random subspace method, you must ensure that you must ensure that the extracted feature subspace has sufficient differences. To this end, we introduce the RSM based on double Bagging sampling algorithm to increase the subspace difference. First, using Bagging sampling from the original sample set, then, using the chi-square statistic based feature selection algorithm to calculate the weight of each feature, and extracting a portion of the more important features.Finally, using Bagging sampling from these features again. Experimental results show that the RSM based on double Bagging can effectively enhance the classifier performance.(2) The RSM adaptive selection algorithm based on OOB. The number of base classifiers and the size of sub-space is random subspace method two important parameters to be determined. Classical subspace methods lack of appropriate guidelines in the base classifiers number and choice of subspace dimension, which need to manually specify, and select the unappropriate parameters will greatly reduce the ensemble classifier performance. Therefore, this article will use the OOB estimation based detection method and max-min linear estimation method to determine the number of base classifiers and the size of sub-space. Experimental results show that the adaptive selection algorithm based OOB can overcome the negative impacts of artificially selected parameters on the random subspace, and improve training integrated classifier training efficiency.(3) RSM selective ensemble algorithm based on FP-Tree. The feature subsets of random subspace method are randomly selected, and the extracted characteristics can not be guaranteed with the necessary ability to distinguish.In this case, it is possible to train the individual classifiers with poor performance, affect the performance of integration. To this end, this paper introduces the idea of selective integration, proposed RSM selective ensemble algorithm based on FP-Tree, which first obtain a streamlined transaction database, and create a FP-Tree to save its contents; then, based on the FP-Tree to obtain the corresponding size of ensemble classifier. Since this method selects only part of the base classifiers in integration can eliminate poor performance of the individual classifiers thus obtained better results than with all base classifiers integration.Finally, the paper also summarizes the three RSM-based improvements.