| As the most common digital media in Internet, digital image with big redundancyis very fit for hiding information. The research result of steganography with digitalimage as the carrier is the most abundant and skilled, which is the most widely used.So the further research for digital image steganalysis is very necessary. This papermainly studies steganalysis on spatial-domain image, the main research results are asfollows:Firstly, by modeling LSB matching based image steganography techniques, weproposed a dectection method based on geometric measures of image histogram. First,LSB matching can be modeled as adding independent additive noise to the image, thiswill lead to image histogram smoothed by a low pass filter. Curvature is the best wayto measure smoothness, and is utilized to evaluate the smoothness of the histogram.Then, the calibration mechanism based secondary steganogtaphy is introduced toreduce the steganalytic difficulty caused by the image variety. SVM are utilized totrain and test the classifiers on large image databases, Experimental results show thatthe proposed method is efficient to detect the LSB matching steganography and hassuperior results compared with the same kind of other algorithms.Secondly, a high-dimensional feature space for steganalysis of LSB matching isproposed based on curvature mode matrix and markov-chain. First, we analysis theimpact of LSB matching to dependences between pixels in nature images, then we getthe curvature mode matrix by nonlinear curvature transformation to the data of imageand model the curvature mode matrix using a markov chain to get the thehigh-dimensional feature space. A feature selection algorithm based on receiveroperating characteristic (ROC) analysis is introduced to obtain the feature subspacewhich suit to steganalysis. Ensemble Classifiers are utilized to train and test theclassifiers on large image databases and experimental results show that the proposedmethod outperforms state-of-the-art techniques. |
PDF Full Text Request