Steganalysis based on Markov Model of Thresholded Prediction-Error Image

Abstract

A steganalysis system based on 2-D Markov chain of thresholded prediction-error image is proposed in this paper. Image pixels are predicted with their neighboring pixels, and the prediction-error image is generated by subtracting the prediction value from the pixel value and then thresholded with a predefined threshold. The empirical transition matrixes of Markov chain along the horizontal, vertical and diagonal directions serve as features for steganalysis. Support vector machines (SVM) are utilized as classifier. The effectiveness of the proposed system has been demonstrated by extensive experimental investigation. The detection rate for Cox et al.'s non-blind spread spectrum (SS) data hiding method, Piva et al.'s blind SS method, and a generic QIM method (as embedding data rate being 0.1 bpp (bits per pixel)) are all above 90% over an image database consisting of approximately 4000 images. For generic LSB method (with various embedding data rates), our steganalysis system achieves a detection rate above 85% as the embedding data rate is 0.1 bpp and above.

Steganalysis by subtractive pixel adjacency matrix

“…It was then further improved to model pixel differences instead of pixel values in [26]. In our paper, we show that there is a great performance benefit in using higher-order models without running into the curse of dimensionality.…”

Steganalysis of LSB Replacement Using Parity-Aware Features

“…Steganalyzers working with features formed as joint or transition probability distributions as features were shown to outperform [27,25,12,11,13] all previously proposed attacks on LSB matching and the content-adaptive HUGO. In summary, it makes perfect sense to expect that the accuracy of the WS detector can be improved as well by considering higher-order statistical constructs from the residual.…”

“…For such embedding operations, the most accurate detectors today are built as classifiers using features obtained as sampled joint distributions (co-occurrence matrices) among neighboring elements of noise residuals [12,11,27,25,13]. These detectors perform equally well for both LSB replacement and LSB matching because features formed from noise residuals are generally blind to pixels' parity.…”

A Survey of Data Mining Techniques for Steganalysis

“…In (Zou et al, 2006), the authors have suggested steganalysis scheme which depends on a 2-D Markov chain thresholded forecast-error image. Image pixels are forecasted with their adjacenting pixels, and the forecast-error image is produced by decreasing the calculated value of the pixel value and then thresholded with a predefined threshold.…”

“…(http://www.datahiding.org), courtesy of Dr. Edward Delp, Purdue University 1000 WAV audio signals files covering different types such as digital speech, on-line broadcast, and music, etc. 19 (Zou et al, 2006) Used 2812 images download from the website of Vision Research Lab, University of California, Santa Barbara. …”