integration de remarques de reviewers

[canny.git] / experiments.tex

diff --git a/experiments.tex b/experiments.tex
index 9af03c672885bd3557429e9dd2cc5f88c3dc0af5..5ba43dfa16ae206dcee8557c46adc9ef457b7645 100644 (file)
--- a/experiments.tex
+++ b/experiments.tex
@@ -1,23 +1,62 @@
-For whole experiments, the whole set of 10,000 images 
+First of all, the whole code of STABYLO can be downloaded
+\footnote{\url{http://members.femto-st.fr/raphael-couturier/stabylo/}}.
+For all the experiments, the whole set of 10,000 images 

 of the BOSS contest~\cite{Boss10} database is taken. 
 In this set, each cover is a $512\times 512$
 grayscale digital image in a RAW format. 
 We restrict experiments to 
 this set of cover images since this paper is more focused on 
 of the BOSS contest~\cite{Boss10} database is taken. 
 In this set, each cover is a $512\times 512$
 grayscale digital image in a RAW format. 
 We restrict experiments to 
 this set of cover images since this paper is more focused on 

-the methodology than benchmarking.    
-Our approach is always compared to Hugo~\cite{DBLP:conf/ih/PevnyFB10}
-and to EAISLSBMR~\cite{Luo:2010:EAI:1824719.1824720}.
-The former is the least detectable information hiding tool in spatial domain 
-and the latter is the work that is the closest to ours, as far as we know. 
-
+the methodology than on benchmarks.    
+
+We use the matrices $\hat{H}$ 
+generated by the integers given
+in Table~\ref{table:matrices:H}
+as introduced in~\cite{FillerJF11}, since these ones have experimentally 
+be proven to have the strongest modification efficiency.
+For instance if the rate between the size of the message and the size of the 
+cover vector
+is 1/4, each number in $\{81, 95, 107, 121\}$ is translated into a binary number 
+and each one constitutes thus a column of $\hat{H}$. 
+
+\begin{table}
+$$
+\begin{array}{|l|l|}
+\hline
+\textrm{Rate} & \textrm{Matrix generators} \\
+\hline
+{1}/{2} & \{71,109\}\\
+\hline
+{1}/{3} & \{95, 101, 121\}\\
+\hline
+{1}/{4} & \{81, 95, 107, 121\}\\
+\hline
+{1}/{5} & \{75, 95, 97, 105, 117\}\\
+\hline
+{1}/{6} & \{73, 83, 95, 103, 109, 123\}\\
+\hline
+{1}/{7} & \{69, 77, 93, 107, 111, 115, 121\}\\
+\hline
+{1}/{8} & \{69, 79, 81, 89, 93, 99, 107, 119\}\\
+\hline
+{1}/{9} & \{69, 79, 81, 89, 93, 99, 107, 119, 125\}\\
+\hline
+\end{array}
+$$
+\caption{Matrix Generator for $\hat{H}$ in STC}\label{table:matrices:H}
+\end{table}

 
 
 
 

-First of all,  in our experiments and with the adaptive scheme, 
-the average size of the message that can be embedded is 16,445 bits.
-Its corresponds to an  average payload of 6.35\%. 
-The two other tools will then be compared with this payload. 
-Sections~\ref{sub:quality} and~\ref{sub:steg} respectively present 
-the quality analysis and the security of our scheme. 
+Our approach is always compared to HUGO, to EAISLSBMR, to WOW and to UNIWARD
+for the two strategies Fixed and Adaptive. 
+For the former one, the payload has been set to 10\%. 
+For the latter one, the Canny parameter $T$ has been set to 3. 
+When $b$ is 7, the average size of the message that can be embedded
+is 16,445 bits, 
+that corresponds to an  average payload of 6.35\%.
+For each cover image the STABYLO's embedding rate with these two parameters 
+is memorized. 
+Next each steganographic scheme is executed to produce the stego content of 
+this cover with respect to this embedding rate.  

 
 
 
 
 
 


@@ -28,14 +67,13 @@ The visual quality of the STABYLO scheme is evaluated in this section.
 For the sake of completeness, three metrics are computed in these experiments: 
 the Peak Signal to Noise Ratio (PSNR), 
 the PSNR-HVS-M family~\cite{psnrhvsm11}, 
 For the sake of completeness, three metrics are computed in these experiments: 
 the Peak Signal to Noise Ratio (PSNR), 
 the PSNR-HVS-M family~\cite{psnrhvsm11}, 

-%the BIQI~\cite{MB10}, 

 and 
 the weighted PSNR (wPSNR)~\cite{DBLP:conf/ih/PereiraVMMP01}.
 The first one is widely used but does not take into
 account the Human Visual System (HVS).
 The other ones have been designed to tackle this problem.
 
 and 
 the weighted PSNR (wPSNR)~\cite{DBLP:conf/ih/PereiraVMMP01}.
 The first one is widely used but does not take into
 account the Human Visual System (HVS).
 The other ones have been designed to tackle this problem.
 

-If we apply them on the running example, 
+If we apply them on the running example with the Adaptive and STC strategies, 

 the PSNR, PSNR-HVS-M, and wPSNR values are respectively equal to 
 68.39, 79.85, and 89.71 for the stego Lena when $b$ is equal to 7.
 If $b$ is 6, these values are respectively equal to 
 the PSNR, PSNR-HVS-M, and wPSNR values are respectively equal to 
 68.39, 79.85, and 89.71 for the stego Lena when $b$ is equal to 7.
 If $b$ is 6, these values are respectively equal to 


@@ -46,30 +84,25 @@ If $b$ is 6, these values are respectively equal to
 
 \begin{table*}
 \begin{center}
 
 \begin{table*}
 \begin{center}

-\begin{tabular}{|c|c|c||c|c|c|c|c|c|}
+\begin{small}
+\setlength{\tabcolsep}{3pt}
+\begin{tabular}{|c|c|c||c|c|c|c|c|c|c|c|c|c|}

 \hline
 \hline

-Schemes & \multicolumn{4}{|c|}{STABYLO} & \multicolumn{2}{|c|}{HUGO}& \multicolumn{2}{|c|}{EAISLSBMR} \\
+Schemes & \multicolumn{4}{|c|}{STABYLO} & \multicolumn{2}{|c|}{HUGO}& \multicolumn{2}{|c|}{EAISLSBMR} & \multicolumn{2}{|c|}{WOW} & \multicolumn{2}{|c|}{UNIWARD}\\

 \hline
 \hline

-Embedding &   Fixed & \multicolumn{3}{|c|}{Adaptive (about 6.35\%)} &  \multicolumn{2}{|c|}{Fixed}& \multicolumn{2}{|c|}{Fixed} \\
+Embedding &   Fixed & \multicolumn{3}{|c|}{Adaptive (about 6.35\%)} &  Fixed &Adaptive & Fixed &Adaptive & Fixed &Adaptive & Fixed &Adaptive \\

 \hline
 \hline

-Rate &   10\% &  + sample &  +STC(7) & +STC(6) &  10\%&6.35\%& 10\%&6.35\%\\ 
+Rate &   10\% &  + sample &  +STC(7) & +STC(6) &  10\%&$\approx$6.35\%& 10\%&$\approx$6.35\%& 10\%&$\approx$6.35\%& 10\%&$\approx$6.35\%\\ 

 \hline
 \hline

-PSNR & 61.86 & 63.48 &  66.55 (\textbf{-0.8\%}) &  63.7  & 64.65 & {67.08} & 60.8 & 62.9\\ 
+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\\ 

 \hline
 \hline

-PSNR-HVS-M & 72.9 & 75.39 & 78.6 (\textbf{-0.8\%}) & 75.5  & 76.67 & {79.23} & 71.8  & 74.3\\ 
-%\hline
-%BIQI & 28.3 & 28.28 & 28.4 & 28.28 & 28.28 & 28.2 & 28.2\\ 
+PSNR-HVS-M & 72.9 & 75.39 & 78.6  & 75.5  & 76.67 & {79.6} & 71.8  & 76.0 &
+76.7 & 80.35 & 77.6 & 81.2 \\

 \hline
 \hline

-wPSNR & 77.47 & 80.59 & 86.43(\textbf{-1.6\%})& 86.28  & 83.03 & {87.8} & 76.7 & 80.6\\ 
+wPSNR & 77.47 & 80.59 & 86.43& 86.28  & 83.03 & {88.6} & 76.7 & 83& 83.8 & 90.4 & 85.2 & 91.9\\ 

 \hline
 \end{tabular}
 \hline
 \end{tabular}

-
-\begin{footnotesize}
-\vspace{2em}
-Variances given in bold font express the quality differences between 
-HUGO and STABYLO with  STC+adaptive parameters.
-\end{footnotesize}
-
+\end{small}

 \end{center}
 \caption{Quality measures of steganography approaches\label{table:quality}}
 \end{table*}
 \end{center}
 \caption{Quality measures of steganography approaches\label{table:quality}}
 \end{table*}


@@ -77,95 +110,106 @@ HUGO and STABYLO with  STC+adaptive parameters.
 
 
 Results are summarized in Table~\ref{table:quality}.
 
 
 Results are summarized in Table~\ref{table:quality}.

+In this table, STC(7) stands for embedding data in the LSB whereas
+in STC(6), data are hidden in the last two  significant bits. 
+
+

 Let us give an interpretation of these experiments.
 Let us give an interpretation of these experiments.

-First of all, the adaptive strategy produces images with lower distortion 
+First of all, the Adaptive strategy produces images with lower distortion 

 than the images resulting from the 10\% fixed strategy.
 Numerical results are indeed always greater for the former strategy than 
 for the latter one.
 than the images resulting from the 10\% fixed strategy.
 Numerical results are indeed always greater for the former strategy than 
 for the latter one.

-These results are not surprising since the adaptive strategy aims at 
-embedding messages whose length is decided according to an higher threshold
+These results are not surprising since the Adaptive strategy aims at 
+embedding messages whose length is decided according to a higher threshold

 into the edge detection.  
 into the edge detection.  

-Let us focus on the quality of HUGO images: with a given fixed 
-embedding rate (10\%), 
-HUGO always produces images whose quality is higher than the STABYLO's one.
-However our approach is always better than EAISLSBMR since this one may modify 
-the two least significant bits.

 
 

-If we combine \emph{adaptive} and \emph{STC} strategies 
-(which leads to an average embedding rate equal to 6.35\%)
-our approach  provides metrics equivalent to those provided by HUGO.
-In this column STC(7) stands for embedding data in the LSB whereas
-in STC(6), data are hidden in the two last significant bits. 

 
 

+If we combine Adaptive and STC strategies 
+the STABYLO scheme  provides images whose quality is higher than 
+the EAISLSBMR's one but lower than the quality of high complexity 
+schemes. Notice that the quality of the less respectful scheme (EAILSBMR) 
+is lower than 6\% than the one of the most one.

 
 
 
 

-The quality variance between HUGO and STABYLO for these parameters
-is given in bold font. It is always close to 1\% which confirms 
-the objective presented in the motivations:
-providing an efficient steganography approach with a lightweight manner.
+% Let us now compare the STABYLO approach with other edge based steganography
+% approaches, namely~\cite{DBLP:journals/eswa/ChenCL10,Chang20101286}.
+% These two schemes focus on increasing the
+% payload while the PSNR is acceptable, but do not 
+% give quality metrics for fixed embedding rates from a large base of images. 

 
 
 
 

-Let us now compare the STABYLO approach with other edge based steganography
-approaches, namely~\cite{DBLP:journals/eswa/ChenCL10,Chang20101286}.
-These two schemes focus on increasing the
-payload while the PSNR is acceptable, but do not 
-give quality metrics for fixed embedding rates from a large base of images. 

 
 
 
 

+\subsection{Steganalysis}\label{sub:steg}

 
 
 
 

-\subsection{Steganalysis}\label{sub:steg}

 
 

+The steganalysis quality of our approach has been evaluated through the % two 
+% AUMP~\cite{Fillatre:2012:ASL:2333143.2333587}
+% and
+Ensemble Classifier~\cite{DBLP:journals/tifs/KodovskyFH12} based steganalyser.
+Its  particularization to spatial domain is 
+considered as state of the art steganalysers.
+Firstly, a space 
+of 686 co-occurrence and Markov features is extracted from the 
+set of cover images and the set of training images. Next a small 
+set of weak classifiers is randomly built,
+each one working on a subspace of all the features.
+The final classifier is constructed by a majority voting 
+between the decisions of these  individual classifiers.

 
 
 
 

-The steganalysis quality of our approach has been evaluated through the two 
-AUMP~\cite{Fillatre:2012:ASL:2333143.2333587}
-and Ensemble Classifier~\cite{DBLP:journals/tifs/KodovskyFH12} based steganalysers.
-Both aim at detecting hidden bits in grayscale natural images and are 
-considered as state of the art steganalysers in the spatial domain~\cite{FK12}.
-The former approach is based on a simplified parametric model of natural images.
-Parameters are firstly estimated and an adaptive Asymptotically Uniformly Most Powerful
-(AUMP) test is designed (theoretically and practically), to check whether
-an image has stego content or not.
-This approach is dedicated to verify whether LSB has been modified or not.
-In the latter, the authors show that the 
-machine learning step, which is often
-implemented as a support vector machine,
-can be favorably executed thanks to an ensemble classifier.
+%The former approach is based on a simplified parametric model of natural images.
+% Parameters are firstly estimated and an adaptive Asymptotically Uniformly Most Powerful
+% (AUMP) test is designed (theoretically and practically), to check whether
+% an image has stego content or not.
+% This approach is dedicated to verify whether LSB has been modified or not.
+% , the authors show that the 
+% machine learning step, which is often
+% implemented as a support vector machine,
+% can be favorably executed thanks to an ensemble classifier.

 
 
 \begin{table*}
 \begin{center}
 
 
 \begin{table*}
 \begin{center}

-%\begin{small}
-\begin{tabular}{|c|c|c|c|c|c|c|c|c|}
-\hline
-Schemes & \multicolumn{4}{|c|}{STABYLO} & \multicolumn{2}{|c|}{HUGO}& \multicolumn{2}{|c|}{EAISLSBMR}\\
+\begin{small}
+\setlength{\tabcolsep}{3pt}
+\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|c|}

 \hline
 \hline

-Embedding & Fixed &   \multicolumn{3}{|c|}{Adaptive (about 6.35\%)}  & \multicolumn{2}{|c|}{Fixed}& \multicolumn{2}{|c|}{Fixed} \\
+Schemes & \multicolumn{4}{|c|}{STABYLO} & \multicolumn{2}{|c|}{HUGO}& \multicolumn{2}{|c|}{EAISLSBMR} &  \multicolumn{2}{|c|}{WOW} &  \multicolumn{2}{|c|}{UNIWARD}\\

 \hline
 \hline

-Rate & 10\% &  + sample &   +STC(7) & +STC(6)   & 10\%& 6.35\%& 10\%& 6.35\%\\ 
+Embedding & Fixed &   \multicolumn{3}{|c|}{Adaptive (about 6.35\%)}  & Fixed & Adapt. & Fixed & Adapt. & Fixed & Adapt. & Fixed & Adapt. \\

 \hline
 \hline

-AUMP & 0.22 & 0.33 & 0.39  &   0.45    &  0.50 &  0.50 & 0.49 & 0.50 \\
+Rate & 10\% &  + sample &   +STC(7) & +STC(6)   & 10\%& $\approx$6.35\%& 10\%& $\approx$6.35\% & 10\%& $\approx$6.35\%& 10\%& $\approx$6.35\%\\ 

 \hline
 \hline

-Ensemble Classifier & 0.35 & 0.44 & 0.47 & 0.47     & 0.48 &  0.49  &  0.43  & 0.46 \\
+%AUMP & 0.22 & 0.33 & 0.39  &   0.45    &  0.50 &  0.50 & 0.49 & 0.50 \\
+%\hline
+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 \\

 
 \hline
 \end{tabular}
 
 \hline
 \end{tabular}

-%\end{small}
+\end{small}

 \end{center}
 \caption{Steganalysing STABYLO\label{table:steganalyse}} 
 \end{table*}
 
 
 \end{center}
 \caption{Steganalysing STABYLO\label{table:steganalyse}} 
 \end{table*}
 
 

-Results are summarized in Table~\ref{table:steganalyse}.
-First of all, STC outperforms the sample strategy for the two steganalysers, as 
+Results of average testing errors 
+are summarized in Table~\ref{table:steganalyse}.
+First of all, STC outperforms the sample strategy %for % the two steganalysers
+ as 

 already noticed in the quality analysis presented in the previous section. 
 already noticed in the quality analysis presented in the previous section. 

-Next, our approach is more easily detectable than HUGO, which
-is the most secure steganographic tool, as far as we know. 
-However by combining \emph{adaptive} and \emph{STC} strategies
-our approach obtains similar results to HUGO ones.
-
+Next, our approach is more easily detectable than HUGO, 
+WOW and UNIWARD which are the most secure steganographic tool, 
+as far as we know. 
+However by combining Adaptive and STC strategies
+our approach obtains similar results than the ones of these schemes.
+
+Compared to EAILSBMR, we obtain better results when the strategy is 
+Adaptive. 

 However due to its 
 However due to its 

-huge number of integration features, it is not lightweight, which justifies 
-in the authors' opinion the consideration of the proposed method.   
+huge number of integration features, it is not lightweight.

 
 

+All these numerical experiments confirm 
+the objective presented in the motivations:
+providing an efficient steganography approach in a lightweight manner.