update of figures

[canny.git] / experiments.tex

diff --git a/experiments.tex b/experiments.tex
index 7afd27cb88c7e0877857d9dceed97dc0955eddec..15fb5c10afca1fafb81bba6ea93ca3fe544301bd 100644 (file)
--- a/experiments.tex
+++ b/experiments.tex
@@ -1,19 +1,22 @@
-For whole experiments, the whole set of 10,000 images 
+First of all, the whole code of STABYLO can be downloaded
+\footnote{\url{http://http://members.femto-st.fr/jf-couchot/en/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.    
+the methodology than on benchmarks.    

 
 We use the matrices $\hat{H}$ 
 generated by the integers given
 
 We use the matrices $\hat{H}$ 
 generated by the integers given

-in table~\ref{table:matrices:H}
+in Table~\ref{table:matrices:H}

 as introduced in~\cite{FillerJF11}, since these ones have experimentally 
 as introduced in~\cite{FillerJF11}, since these ones have experimentally 

-be proven to have the best modification efficiency.
-For instance if the rate between the size of message and the size of the host is 
-1/4, each number in $\{81, 95, 107, 121\}$ is translated into a binary number 
-and each one consitutes thus an column of $\hat{H}$. 
+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{table}
 $$


@@ -39,115 +42,100 @@ $$
 \hline
 \end{array}
 $$
 \hline
 \end{array}
 $$

-\caption{Matrix Generator for $\hat{H}$ in STC}\label{table:matrices:H}
+\caption{Matrix Generator for $\hat{H}$ in STC.}\label{table:matrices:H}

 \end{table}
 
 
 \end{table}
 
 

-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. 
-
-
-
-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.  

 
 
 
  
 
 
 
 
  
 

-\subsection{Image quality}\label{sub:quality}
-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}, 
-%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.
-
-If we apply them on the running example, 
-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 
-65.43, 77.2, and 89.35.
+% \subsection{Image quality}\label{sub:quality}
+% 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}, 
+% 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 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 
+% 65.43, 77.2, and 89.35.

 
 
 
 
 
 

-\begin{table*}
-\begin{center}
-\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} \\
-\hline
-Embedding &   Fixed & \multicolumn{3}{|c|}{Adaptive (about 6.35\%)} &  \multicolumn{2}{|c|}{Fixed}& \multicolumn{2}{|c|}{Fixed} \\
-\hline
-Rate &   10\% &  + sample &  +STC(7) & +STC(6) &  10\%&6.35\%& 10\%&6.35\%\\ 
-\hline
-PSNR & 61.86 & 63.48 &  66.55 (\textbf{-0.8\%}) &  63.7  & 64.65 & {67.08} & 60.8 & 62.9\\ 
-\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\\ 
-\hline
-wPSNR & 77.47 & 80.59 & 86.43(\textbf{-1.6\%})& 86.28  & 83.03 & {87.8} & 76.7 & 80.6\\ 
-\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{center}
-\caption{Quality measures of steganography approaches\label{table:quality}}
-\end{table*}
+% \begin{table*}
+% \begin{center}
+% \begin{small}
+% \setlength{\tabcolsep}{3pt}
+% \begin{tabular}{|c|c|c||c|c|c|c|c|c|c|c|c|c|}
+% \hline
+% Schemes & \multicolumn{4}{|c|}{STABYLO} & \multicolumn{2}{|c|}{HUGO}& \multicolumn{2}{|c|}{EAISLSBMR} & \multicolumn{2}{|c|}{WOW} & \multicolumn{2}{|c|}{UNIWARD}\\
+% \hline
+% Embedding &   Fixed & \multicolumn{3}{|c|}{Adaptive (about 6.35\%)} &  Fixed &Adaptive & Fixed &Adaptive & Fixed &Adaptive & Fixed &Adaptive \\
+% \hline
+% Rate &   10\% &  + sample &  +STC(7) & +STC(6) &  10\%&$\approx$6.35\%& 10\%&$\approx$6.35\%& 10\%&$\approx$6.35\%& 10\%&$\approx$6.35\%\\ 
+% \hline
+% 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
+% 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
+% 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}
+% \end{small}
+% \end{center}
+% \caption{Quality measures of steganography approaches\label{table:quality}}
+% \end{table*}

 
 
 
 
 
 

-Results are summarized in Table~\ref{table:quality}.
-Let us give an interpretation of these experiments.
-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.
-These results are not surprising since the adaptive strategy aims at 
-embedding messages whose length is decided according to an higher threshold
-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.
+% 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. 

 
 

-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. 

 
 

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

 
 
 
 

-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.
+% 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.

 
 
 
 

-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. 

 
 
 
 
 
 


@@ -156,54 +144,94 @@ give quality metrics for fixed embedding rates from a large base of images.
 
 
 
 
 
 

-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 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.
+\JFC{Features that are embedded into this steganalysis process 
+are CCPEV and SPAM features as described 
+in~\cite{DBLP:dblp_conf/mediaforensics/KodovskyPF10}.
+They  are 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 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}
 \end{center}

-\caption{Steganalysing STABYLO\label{table:steganalyse}} 
+\caption{Steganalysing STABYLO\label{table:steganalyse}.} 

 \end{table*}
 
 
 \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.
-
-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.   
-
+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 similar
+results when the strategy is 
+Adaptive. 
+However due to its 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
+for small payload.
+
+\RC{In Figure~\ref{fig:error}, 
+Ensemble Classifier has been used with all the previous 
+steganographic schemes with 4 different payloads.
+It can be observed that face to high values of payload, 
+STABYLO is definitely not secure enough.
+However thanks to an efficient very low-complexity (Fig.\ref{fig:compared}), 
+we argue that the user should embed tiny messages in many images 
+than a larger message in only one image.
+\begin{figure}
+\begin{center}
+\includegraphics[scale=0.5]{error}
+\end{center}
+\caption{Testing error obtained by Ensemble classifier with 
+WOW/UNIWARD, HUGO, and STABYLO w.r.t. payload.}
+\label{fig:error} 
+\end{figure}
+}