FEATURE SELECTION FOR BLIND IMAGE STEGANALYSIS

Abstract

Blind Steganalysis is the art and science of detecting, extracting or destroying presence of hidden messages in innocent looking cover mediums such as digital images, videos, audios and texts for a known or unknown steganography algorithm. This research focuses on JPEG images because of its ubiquitous nature and low bandwidth requirements for storage and transmission. newlineBlind image steganalysis is generally implemented in two steps during first step statistical features which are sensitive to message embedding are extracted and second step involves a classifier model that helps distinguish a cover image (original image) from a stego image (image with secret message). The improvement in performance of steganalyzers has been achieved over the years by increasing the feature space which causes Curse of dimensionality . One of the parameters that determine the complexity of classifier model is dimensionality of feature space. The high dimensional feature space increases the computational complexity and some of the features may be redundant or irrelevant which could worsen the performance of the classifier. newlineThe purpose of this thesis is to overcome Curse of dimensionality by selecting a subset of relevant features that can efficiently classify images as stego or cover and reduce computational complexity of the classifier. To achieve this three hybrid models have been proposed. Each of the models consists of two phases. Phase I is an ensemble of univariate (t-test) and multivariate (Multiple Regression) filter feature selection algorithms which selects relevant features to differentiate between stego and cover images and forwards the selected feature subset to second phase. newlineIn Phase II, variations of three different heuristic wrapper approaches are suggested to get optimal classification accuracy at reduced computational cost by working on significant feature space selected during Phase I. First wrapper approach varies Discrete Particle Swarm Optimization (DPSO) to overcome the drawback of Global Best PSO getting trapped in local optima. It combines Global Best PSO with Local Best PSO and applies Hope/Rehope concept. The second wrapper approach introduces methods of dynamic adaptation to improve convergence rate and a probabilistic approach to reduce computational complexity of Discrete Firefly Algorithm (DFA). The third wrapper approach applies a modified function for discretization of population of birds and varies cognitive, social and flight parameters to improve performance of Bird Swarm Algorithm (BSA). newlineIn this study, Support Vector Machine (SVM) classifier with Radial Basis Function (RBF) kernel and 10 fold cross validation is used to evaluate the effectiveness of proposed algorithms and also to compute the fitness function of heuristic approaches. The proposed models are tested on four different steganography algorithms namely nsF5, Perturbed Quantization, Outguess and Steghide with two sets of feature vectors Cartesian Calibrated Pevný (CCPEV) and Subtractive Pixel Adjacency Matrix (SPAM). All three models outperform conventional filter and wrapper feature selection algorithms. Statistical significant test performed with two tailed t-test by setting the significance level at less than 0.01 demonstrates the superiority of the proposed approaches over well-known wrapper approaches. newlineFinally, an analysis of low dimensional feature vectors identified by all three hybrid models is made to find out contribution of each type of statistical feature in detecting the hidden message. The effectiveness of selected features is also experimentally tested for a new steganography algorithm JP Hide and Seek (JPHS). The results show remarkable improvement in classification accuracy thereby indicating that the proposed hybrid approaches are highly effective in detecting images as stego or cover even from an unknown steganography algorithm. newline newline