1 For whole experiments, the whole set of 10,000 images
2 of the BOSS contest~\cite{Boss10} database is taken.
3 In this set, each cover is a $512\times 512$
4 grayscale digital image in a RAW format.
5 We restrict experiments to
6 this set of cover images since this paper is more focused on
7 the methodology than benchmarking.
9 We use the matrices $\hat{H}$
10 generated by the integers given
11 in table~\ref{table:matrices:H}
12 as introduced in~\cite{FillerJF11}, since these ones have experimentally
13 be proven to have the best modification efficiency.
14 For instance if the rate between the size of message and the size of the host is
15 1/4, each number in $\{81, 95, 107, 121\}$ is translated into a binary number
16 and each one consitutes thus an column of $\hat{H}$.
22 \textrm{Rate} & \textrm{Matrix generators} \\
24 {1}/{2} & \{71,109\}\\
26 {1}/{3} & \{95, 101, 121\}\\
28 {1}/{4} & \{81, 95, 107, 121\}\\
30 {1}/{5} & \{75, 95, 97, 105, 117\}\\
32 {1}/{6} & \{73, 83, 95, 103, 109, 123\}\\
34 {1}/{7} & \{69, 77, 93, 107, 111, 115, 121\}\\
36 {1}/{8} & \{69, 79, 81, 89, 93, 99, 107, 119\}\\
38 {1}/{9} & \{69, 79, 81, 89, 93, 99, 107, 119, 125\}\\
42 \caption{Matrix Generator for $\hat{H}$ in STC}\label{table:matrices:H}
46 Our approach is always compared to Hugo~\cite{DBLP:conf/ih/PevnyFB10}
47 and to EAISLSBMR~\cite{Luo:2010:EAI:1824719.1824720}.
48 The former is the least detectable information hiding tool in spatial domain
49 and the latter is the work that is the closest to ours, as far as we know.
53 First of all, in our experiments and with the adaptive scheme,
54 the average size of the message that can be embedded is 16,445 bits.
55 Its corresponds to an average payload of 6.35\%.
56 The two other tools will then be compared with this payload.
57 Sections~\ref{sub:quality} and~\ref{sub:steg} respectively present
58 the quality analysis and the security of our scheme.
64 \subsection{Image quality}\label{sub:quality}
65 The visual quality of the STABYLO scheme is evaluated in this section.
66 For the sake of completeness, three metrics are computed in these experiments:
67 the Peak Signal to Noise Ratio (PSNR),
68 the PSNR-HVS-M family~\cite{psnrhvsm11},
69 %the BIQI~\cite{MB10},
71 the weighted PSNR (wPSNR)~\cite{DBLP:conf/ih/PereiraVMMP01}.
72 The first one is widely used but does not take into
73 account the Human Visual System (HVS).
74 The other ones have been designed to tackle this problem.
76 If we apply them on the running example,
77 the PSNR, PSNR-HVS-M, and wPSNR values are respectively equal to
78 68.39, 79.85, and 89.71 for the stego Lena when $b$ is equal to 7.
79 If $b$ is 6, these values are respectively equal to
80 65.43, 77.2, and 89.35.
87 \begin{tabular}{|c|c|c||c|c|c|c|c|c|}
89 Schemes & \multicolumn{4}{|c|}{STABYLO} & \multicolumn{2}{|c|}{HUGO}& \multicolumn{2}{|c|}{EAISLSBMR} \\
91 Embedding & Fixed & \multicolumn{3}{|c|}{Adaptive (about 6.35\%)} & \multicolumn{2}{|c|}{Fixed}& \multicolumn{2}{|c|}{Fixed} \\
93 Rate & 10\% & + sample & +STC(7) & +STC(6) & 10\%&6.35\%& 10\%&6.35\%\\
95 PSNR & 61.86 & 63.48 & 66.55 (\textbf{-0.8\%}) & 63.7 & 64.65 & {67.08} & 60.8 & 62.9\\
97 PSNR-HVS-M & 72.9 & 75.39 & 78.6 (\textbf{-0.8\%}) & 75.5 & 76.67 & {79.23} & 71.8 & 74.3\\
99 %BIQI & 28.3 & 28.28 & 28.4 & 28.28 & 28.28 & 28.2 & 28.2\\
101 wPSNR & 77.47 & 80.59 & 86.43(\textbf{-1.6\%})& 86.28 & 83.03 & {87.8} & 76.7 & 80.6\\
107 Variances given in bold font express the quality differences between
108 HUGO and STABYLO with STC+adaptive parameters.
112 \caption{Quality measures of steganography approaches\label{table:quality}}
117 Results are summarized in Table~\ref{table:quality}.
118 Let us give an interpretation of these experiments.
119 First of all, the adaptive strategy produces images with lower distortion
120 than the images resulting from the 10\% fixed strategy.
121 Numerical results are indeed always greater for the former strategy than
123 These results are not surprising since the adaptive strategy aims at
124 embedding messages whose length is decided according to an higher threshold
125 into the edge detection.
126 Let us focus on the quality of HUGO images: with a given fixed
127 embedding rate (10\%),
128 HUGO always produces images whose quality is higher than the STABYLO's one.
129 However our approach is always better than EAISLSBMR since this one may modify
130 the two least significant bits.
132 If we combine \emph{adaptive} and \emph{STC} strategies
133 (which leads to an average embedding rate equal to 6.35\%)
134 our approach provides metrics equivalent to those provided by HUGO.
135 In this column STC(7) stands for embedding data in the LSB whereas
136 in STC(6), data are hidden in the two last significant bits.
140 The quality variance between HUGO and STABYLO for these parameters
141 is given in bold font. It is always close to 1\% which confirms
142 the objective presented in the motivations:
143 providing an efficient steganography approach with a lightweight manner.
146 Let us now compare the STABYLO approach with other edge based steganography
147 approaches, namely~\cite{DBLP:journals/eswa/ChenCL10,Chang20101286}.
148 These two schemes focus on increasing the
149 payload while the PSNR is acceptable, but do not
150 give quality metrics for fixed embedding rates from a large base of images.
155 \subsection{Steganalysis}\label{sub:steg}
159 The steganalysis quality of our approach has been evaluated through the two
160 AUMP~\cite{Fillatre:2012:ASL:2333143.2333587}
161 and Ensemble Classifier~\cite{DBLP:journals/tifs/KodovskyFH12} based steganalysers.
162 Both aim at detecting hidden bits in grayscale natural images and are
163 considered as state of the art steganalysers in the spatial domain~\cite{FK12}.
164 The former approach is based on a simplified parametric model of natural images.
165 Parameters are firstly estimated and an adaptive Asymptotically Uniformly Most Powerful
166 (AUMP) test is designed (theoretically and practically), to check whether
167 an image has stego content or not.
168 This approach is dedicated to verify whether LSB has been modified or not.
169 In the latter, the authors show that the
170 machine learning step, which is often
171 implemented as a support vector machine,
172 can be favorably executed thanks to an ensemble classifier.
178 \begin{tabular}{|c|c|c|c|c|c|c|c|c|}
180 Schemes & \multicolumn{4}{|c|}{STABYLO} & \multicolumn{2}{|c|}{HUGO}& \multicolumn{2}{|c|}{EAISLSBMR}\\
182 Embedding & Fixed & \multicolumn{3}{|c|}{Adaptive (about 6.35\%)} & \multicolumn{2}{|c|}{Fixed}& \multicolumn{2}{|c|}{Fixed} \\
184 Rate & 10\% & + sample & +STC(7) & +STC(6) & 10\%& 6.35\%& 10\%& 6.35\%\\
186 AUMP & 0.22 & 0.33 & 0.39 & 0.45 & 0.50 & 0.50 & 0.49 & 0.50 \\
188 Ensemble Classifier & 0.35 & 0.44 & 0.47 & 0.47 & 0.48 & 0.49 & 0.43 & 0.46 \\
194 \caption{Steganalysing STABYLO\label{table:steganalyse}}
198 Results are summarized in Table~\ref{table:steganalyse}.
199 First of all, STC outperforms the sample strategy for the two steganalysers, as
200 already noticed in the quality analysis presented in the previous section.
201 Next, our approach is more easily detectable than HUGO, which
202 is the most secure steganographic tool, as far as we know.
203 However by combining \emph{adaptive} and \emph{STC} strategies
204 our approach obtains similar results to HUGO ones.
207 huge number of integration features, it is not lightweight, which justifies
208 in the authors' opinion the consideration of the proposed method.