摘要：It is hard to detect LSB matching steganography, especially in low embedding rate. However there are still some chances to attack it since the data embedding alters the dependences between neighboring pixels. Based on this fact, this paper proposes a novel steganalysis method by modeling the dependences. The neighboring pixels are divided into three groups: horizontal, vertical, and diagonal. Then, the prediction errors of the central pixel are calculated by each group respectively. Finally, the empirical probability matrices among these prediction errors are computed and used as features for steganalysis. Experimental results show the proposed method has better performance than SPAM scheme, which currently is the most sensitive detector for LSB Matching. Furthermore, combined with the SPAM features, the method achieves the best accuracy.
