earticle

영어

Feature based image Steganalysis demands the best feature model for accurate steganalysis. The extracted feature model includes the components in DCT features of JPEG image. Existing research in this field show extraction of different types of image features that show slightly improved classification accuracies. Though few recent methods of image steganalysis involve extracting all possible features of the image, they suffer dimensionality problem. The dataset used in our research include raw images from the BOSS database. The original dimension of the feature set extracted has 8726 features from 2000 images. While a larger feature set is expected to have all important information about the steganographic changes, it affects the classifier accuracy due to redundancy. To overcome the curse of dimensionality, we intend to introduce an unsupervised optimization technique before classification. The individual classifiers implemented are SVM and MLP and the fusion techniques implemented to combine these classifiers are Bayes, Dempster Schafer and Decision Template schemes. The performances of classifiers are analyzed for optimization based on Euclidean distance measure and Mahalanobis distance measure. Comparing individual classifiers, it has been found that SVM classifier outperforms MLP classifier for both Euclidean distance measure and Mahalanobis distance measure. Among the fusion schemes, the accuracy of Bayes fusion scheme proves to be best compared to Decision template and Dempster Schafer schemes. Also, the best possible classification accuracy has been obtained for Euclidean distance based optimization followed by Bayes fusion classifier scheme. The classification accuracies obtained in our research are better compared to existing methods.