A more secure steganography based on adaptive pixel-value ...

Multimed Tools ApplIf gi ′ or gi+1 ′ is out of the range [0,255], then the candidate embedding unit ismarked as abandoned one. Note that such unused units are few in most naturalimages and are very easy to detect.It can be proven that the new absolute difference |d ′ |=|gi+1 ′ − g′ i| in the stegoimage will fall into the same region R k as the difference |d| =|g i+1 − g i | in thecover image. So in data extraction, if |d ′ |∈R k , the embedded value can be extractedcorrectly by b =|d ′ |−l k .In [25], the authors shows that a fixed region R i employed in the original PVDscheme (Step 2) will introduce some unusual steps in the histogram of pixel valuedifferences (PVD histogram) for all the embedding units, which can be used toexpose the presence of hidden message and further to estimate the size of hiddenbits. To make the <strong>steganography</strong> immune to PVD histogram <strong>based</strong> analysis, theauthors employ random regions instead of the fixed regions when data hiding. Theexperimental results in [25] show that the new approach can effectively eliminatethose undesired steps.To improve the embedding capacity, a new PVD-<strong>based</strong> <strong>steganography</strong> combinedwith LSB technique is proposed by Wu et al. [23]. In this method, Steps 1 and 2 arethe same as it did in the original PVD scheme. After that, if R k belongs to higherlevel, namely |d| > 15, then embeds secret bits using the original method. Otherwiseif |d| belongs to lower level, namely |d| ≤15, embeds three bits into g i and g i+1 using3-LSB replacement, respectively. Assume that the pixel values after data hiding aredenoted as gi ′ and g′ i+1 ,thencalculate|d′ |=|gi+1 ′ − g′ i |, if the new difference |d′ | is notbelongs to lower level, then readjust them by(g′i , g ′ i+1)={ (gi − 8, g i+1 + 8 ) g ′ i ≥ g′ i+1(gi + 8, g i+1 − 8 ) g ′ i < g′ i+1In a latest work [24], it divides the pixel differences into three levels: lowerlevel,middle-level and high-level. Similar to Wu’s approach [23], k secret bits areembedded into g i and g i+1 , respectively, where k = l, m, h, if|d| belongs to lowerlevel,middle-level and high-level, respectively. After that the method applies themodified LSB approach [1, 5] to the resulting pixels pair and gets the new pixel pair(gi ′, g′ i+1 ). If the new difference |d′ |=|gi+1 ′ − g′ i| belongs to different levels, readjustthem into the following forms:(g′i , g i+1 ′ ± 2k) (or g′i ± 2 k , g i+1)′Finally selects the better choice of the new values gi ′, g′ i+1which satisfies theconditions that the new difference |d ′ | and the original one belong to the samelevel and the value of |gi ′ − g i|+|gi+1 ′ − g i+1| is the smallest among the candidates.Compared with [23], the new method can provide stego image with larger embeddingcapacity as well as higher objective quality.2.2 Properties of PVD-<strong>based</strong> <strong>steganography</strong>As described in Subsection 2.1, the main idea of the existing PVD-<strong>based</strong> approachesis that they first divide a cover image into non-overlapping embedding units withtwo consecutive pixels in a raster scanning manner, and then deal with the embeddingunits separately in a pseudo-random order. For a given embedding unit, the