1 First of all, the whole code of STABYLO can be downloaded
2 \footnote{\url{http://http://members.femto-st.fr/jf-couchot/en/stabylo}}.
3 For all the experiments, the whole set of 10,000 images
4 of the BOSS contest~\cite{Boss10} database is taken.
5 In this set, each cover is a $512\times 512$
6 grayscale digital image in a RAW format.
7 We restrict experiments to
8 this set of cover images since this paper is more focused on
9 the methodology than on benchmarks.
11 We use the matrices $\hat{H}$
12 generated by the integers given
13 in Table~\ref{table:matrices:H}
14 as introduced in~\cite{FillerJF11}, since these ones have experimentally
15 be proven to have the strongest modification efficiency.
16 For instance if the rate between the size of the message and the size of the
18 is 1/4, each number in $\{81, 95, 107, 121\}$ is translated into a binary number
19 and each one constitutes thus a column of $\hat{H}$.
25 \textrm{Rate} & \textrm{Matrix generators} \\
27 {1}/{2} & \{71,109\}\\
29 {1}/{3} & \{95, 101, 121\}\\
31 {1}/{4} & \{81, 95, 107, 121\}\\
33 {1}/{5} & \{75, 95, 97, 105, 117\}\\
35 {1}/{6} & \{73, 83, 95, 103, 109, 123\}\\
37 {1}/{7} & \{69, 77, 93, 107, 111, 115, 121\}\\
39 {1}/{8} & \{69, 79, 81, 89, 93, 99, 107, 119\}\\
41 {1}/{9} & \{69, 79, 81, 89, 93, 99, 107, 119, 125\}\\
45 \caption{Matrix Generator for $\hat{H}$ in STC}\label{table:matrices:H}
49 Our approach is always compared to HUGO, to EAISLSBMR, to WOW and to UNIWARD
50 for the two strategies Fixed and Adaptive.
51 For the former one, the payload has been set to 10\%.
52 For the latter one, the Canny parameter $T$ has been set to 3.
53 When $b$ is 7, the average size of the message that can be embedded
55 that corresponds to an average payload of 6.35\%.
56 For each cover image the STABYLO's embedding rate with these two parameters
58 Next each steganographic scheme is executed to produce the stego content of
59 this cover with respect to this embedding rate.
65 % \subsection{Image quality}\label{sub:quality}
66 % The visual quality of the STABYLO scheme is evaluated in this section.
67 % For the sake of completeness, three metrics are computed in these experiments:
68 % the Peak Signal to Noise Ratio (PSNR),
69 % the PSNR-HVS-M family~\cite{psnrhvsm11},
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 with the Adaptive and STC strategies,
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.
88 % \setlength{\tabcolsep}{3pt}
89 % \begin{tabular}{|c|c|c||c|c|c|c|c|c|c|c|c|c|}
91 % Schemes & \multicolumn{4}{|c|}{STABYLO} & \multicolumn{2}{|c|}{HUGO}& \multicolumn{2}{|c|}{EAISLSBMR} & \multicolumn{2}{|c|}{WOW} & \multicolumn{2}{|c|}{UNIWARD}\\
93 % Embedding & Fixed & \multicolumn{3}{|c|}{Adaptive (about 6.35\%)} & Fixed &Adaptive & Fixed &Adaptive & Fixed &Adaptive & Fixed &Adaptive \\
95 % Rate & 10\% & + sample & +STC(7) & +STC(6) & 10\%&$\approx$6.35\%& 10\%&$\approx$6.35\%& 10\%&$\approx$6.35\%& 10\%&$\approx$6.35\%\\
97 % PSNR & 61.86 & 63.48 & 66.55 & 63.7 & 64.65 & {67.08} & 60.8 & 62.9&65.9 & 68.3 & 65.8 & 69.2\\
99 % PSNR-HVS-M & 72.9 & 75.39 & 78.6 & 75.5 & 76.67 & {79.6} & 71.8 & 76.0 &
100 % 76.7 & 80.35 & 77.6 & 81.2 \\
102 % wPSNR & 77.47 & 80.59 & 86.43& 86.28 & 83.03 & {88.6} & 76.7 & 83& 83.8 & 90.4 & 85.2 & 91.9\\
107 % \caption{Quality measures of steganography approaches\label{table:quality}}
112 % Results are summarized in Table~\ref{table:quality}.
113 % In this table, STC(7) stands for embedding data in the LSB whereas
114 % in STC(6), data are hidden in the last two significant bits.
117 % Let us give an interpretation of these experiments.
118 % First of all, the Adaptive strategy produces images with lower distortion
119 % than the images resulting from the 10\% fixed strategy.
120 % Numerical results are indeed always greater for the former strategy than
121 % for the latter one.
122 % These results are not surprising since the Adaptive strategy aims at
123 % embedding messages whose length is decided according to a higher threshold
124 % into the edge detection.
127 % If we combine Adaptive and STC strategies
128 % the STABYLO scheme provides images whose quality is higher than
129 % the EAISLSBMR's one but lower than the quality of high complexity
130 % schemes. Notice that the quality of the less respectful scheme (EAILSBMR)
131 % is lower than 6\% than the one of the most one.
134 % % Let us now compare the STABYLO approach with other edge based steganography
135 % % approaches, namely~\cite{DBLP:journals/eswa/ChenCL10,Chang20101286}.
136 % % These two schemes focus on increasing the
137 % % payload while the PSNR is acceptable, but do not
138 % % give quality metrics for fixed embedding rates from a large base of images.
143 \subsection{Steganalysis}\label{sub:steg}
147 The steganalysis quality of our approach has been evaluated through the % two
148 % AUMP~\cite{Fillatre:2012:ASL:2333143.2333587}
150 Ensemble Classifier~\cite{DBLP:journals/tifs/KodovskyFH12} based steganalyser.
151 Its particularization to spatial domain is
152 considered as state of the art steganalysers.
153 \JFC{Features that are embedded into this steganalysis process
154 are CCPEV and SPAM features as described
155 in~\cite{DBLP:dblp_conf/mediaforensics/KodovskyPF10}.
156 These latter are extracted from the
157 set of cover images and the set of training images.}
159 set of weak classifiers is randomly built,
160 each one working on a subspace of all the features.
161 The final classifier is constructed by a majority voting
162 between the decisions of these individual classifiers.
165 %The former approach is based on a simplified parametric model of natural images.
166 % Parameters are firstly estimated and an adaptive Asymptotically Uniformly Most Powerful
167 % (AUMP) test is designed (theoretically and practically), to check whether
168 % an image has stego content or not.
169 % This approach is dedicated to verify whether LSB has been modified or not.
170 % , the authors show that the
171 % machine learning step, which is often
172 % implemented as a support vector machine,
173 % can be favorably executed thanks to an ensemble classifier.
179 \setlength{\tabcolsep}{3pt}
180 \begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|c|}
182 Schemes & \multicolumn{4}{|c|}{STABYLO} & \multicolumn{2}{|c|}{HUGO}& \multicolumn{2}{|c|}{EAISLSBMR} & \multicolumn{2}{|c|}{WOW} & \multicolumn{2}{|c|}{UNIWARD}\\
184 Embedding & Fixed & \multicolumn{3}{|c|}{Adaptive (about 6.35\%)} & Fixed & Adapt. & Fixed & Adapt. & Fixed & Adapt. & Fixed & Adapt. \\
186 Rate & 10\% & + sample & +STC(7) & +STC(6) & 10\%& $\approx$6.35\%& 10\%& $\approx$6.35\% & 10\%& $\approx$6.35\%& 10\%& $\approx$6.35\%\\
188 %AUMP & 0.22 & 0.33 & 0.39 & 0.45 & 0.50 & 0.50 & 0.49 & 0.50 \\
190 Ensemble Classifier & 0.35 & 0.44 & 0.47 & 0.47 & 0.48 & 0.49 & 0.43 & 0.47 & 0.48 & 0.49 & 0.46 & 0.49 \\
196 \caption{Steganalysing STABYLO\label{table:steganalyse}}
200 Results of average testing errors
201 are summarized in Table~\ref{table:steganalyse}.
202 First of all, STC outperforms the sample strategy %for % the two steganalysers
204 already noticed in the quality analysis presented in the previous section.
205 Next, our approach is more easily detectable than HUGO,
206 WOW and UNIWARD which are the most secure steganographic tool,
208 However by combining Adaptive and STC strategies
209 our approach obtains similar results than the ones of these schemes.
211 Compared to EAILSBMR, we obtain similar
212 results when the strategy is
214 However due to its huge number of integration features, it is not lightweight.
216 All these numerical experiments confirm
217 the objective presented in the motivations:
218 providing an efficient steganography approach in a lightweight manner.
220 \RC{In Figure~\ref{fig:error}, Ensemble Classifier has been used with all the previsou steganalizers with 3 different payloads. It can be observed that with important payload, STABYLO is not efficient, but as mentionned its complexity is far more simple compared to other tools.\\
223 \includegraphics[scale=0.5]{error}
225 \caption{Error obtained by Ensemble classifier with WOW/UNIWARD, HUGO, and STABYLO and different paylaods.}