integration de remarques de reviewers

[canny.git] / experiments.tex

diff --git a/experiments.tex b/experiments.tex
index c92062c01027a8f4eed07bcf6fcf9b4e835d9192..5ba43dfa16ae206dcee8557c46adc9ef457b7645 100644 (file)
--- a/experiments.tex
+++ b/experiments.tex
@@ -1,152 +1,215 @@
-For whole experiments, a set of 500 images is randomly extracted 
-from the database taken from the BOSS contest~\cite{Boss10}. 
+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 $512\times 512$
 In this set, each cover is a $512\times 512$

-grayscale digital image.
+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 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 $\hat{H}$. 

 
 

-
-\subsection{Adaptive Embedding Rate} 
-
-Two strategies have been developed in our scheme, depending on the embedding rate that is either \emph{adaptive} or \emph{fixed}.
-
-In the former the embedding rate depends on the number of edge pixels.
-The higher it is, the larger is the message length that can be inserted.
-Practically, a set of edge pixels is computed according to the 
-Canny algorithm with an high threshold.
-The message length is thus defined to be the half of this set cardinality.
-In this strategy, two methods are thus applied to extract bits that 
-are modified. The first one is a direct application of the STC algorithm.
-This method is further referred as \emph{adaptive+STC}.
-The second one randomly choose the subset of pixels to modify by 
-applying the BBS PRNG again. This method is denoted \emph{adaptive+sample}.
-Notice that the rate between 
-available bits  and bit message length is always equal to 2.
-This constraint is indeed induced by the fact that the efficiency 
-of the STC algorithm is unsatisfactory under that threshold.
-On our experiments and with the adaptive scheme, 
-the average size of the message that can be embedded is 16445.
-Its corresponds to an  average payload of 6.35\%. 
+\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}

 
 
 
 

+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.  

 
 
 
 

-In the latter, the embedding rate is defined as a percentage between the 
-number of the modified pixels and the length of the bit message.
-This is the classical approach adopted in steganography.
-Practically, the Canny algorithm generates a 
-a set of edge pixels with threshold that is decreasing until its cardinality
-is sufficient. If the set cardinality is more than twice larger than the 
-bit message length, a STC step is again applied.
-Otherwise, pixels are again randomly chosen with BBS.

 
  
 
 
  
 

-\subsection{Image Quality}
+\subsection{Image quality}\label{sub:quality}

 The visual quality of the STABYLO scheme is evaluated in this section.
 The visual quality of the STABYLO scheme is evaluated in this section.

-For the sake of completeness, four metrics are computed in these experiments: 
+For the sake of completeness, three metrics are computed in these experiments: 

 the Peak Signal to Noise Ratio (PSNR), 
 the Peak Signal to Noise Ratio (PSNR), 

-the PSNR-HVS-M family~\cite{PSECAL07,psnrhvsm11} , 
-the BIQI~\cite{MB10,biqi11}, and 
+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 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 last ones have been designed to tackle this problem.
+The other ones have been designed to tackle this problem.

 
 

-\begin{table}
+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{center}

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

 \hline
 \hline

-Embedding &   \multicolumn{2}{|c||}{Adaptive} & Fixed & 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 &   + STC &  + sample & 10\% & 10\%\\ 
+Embedding &   Fixed & \multicolumn{3}{|c|}{Adaptive (about 6.35\%)} &  Fixed &Adaptive & Fixed &Adaptive & Fixed &Adaptive & Fixed &Adaptive \\

 \hline
 \hline

-PSNR &  66.55 & 63.48  & 61.86  & 64.65   \\ 
+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-HVS-M & 78.6 & 75.39 & 72.9 & 76.67\\ 
+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

-BIQI & 28.3 & 28.28 & 28.4 & 28.28\\ 
+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 & 86.43& 80.59 & 77.47& 83.03\\ 
+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}

+\end{small}

 \end{center}
 \end{center}

-\caption{Quality Measures of Steganography Approaches\label{table:quality}} 
-\end{table}
+\caption{Quality measures of steganography approaches\label{table:quality}}
+\end{table*}
+
+
+
+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 
-than the one of images resulting from the 10\% fixed strategy.
+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 
 Numerical results are indeed always greater for the former strategy than 

-for the latter, except for the BIQI metrics where differences are not relevant.
-These results are not surprising since the adaptive strategy aims at 
+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.  
 embedding messages whose length is decided according to a 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 nevertheless provides better results with the strategy 
-\emph{adaptive+STC} in a lightweight manner, as motivated in the introduction.      

 
 
 
 

-Let us now compare the STABYLO approach with other edge based steganography
-schemes with respect to the image quality.
-First of all, the Edge Adaptive
-scheme detailed in~\cite{Luo:2010:EAI:1824719.1824720} 
-executed with a 10\% embedding rate 
-has the same PSNR but a lower wPSNR than ours:
-these two metrics are respectively equal to 61.9 and 68.9. 
-Next both the approaches~\cite{DBLP:journals/eswa/ChenCL10,Chang20101286}
-focus on increasing the payload while the PSNR is acceptable, but do not 
-give quality metrics for fixed embedding rate from a large base of images. 
-Our approach outperforms the former thanks to the introduction of the STC 
-algorithm.
+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. 

 
 
 
 

-\subsection{Steganalysis}

 
 
 
 

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

 
 

-The 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 aims at detecting hidden bits in grayscale natural images and are 
-considered as the state of the art of steganalysers in 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.  
-In the latter, the authors show that the 
-machine learning step, (which is often
-implemented as support vector machine)
-can be favorably executed thanks to an Ensemble 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 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.

 
 

-\begin{table}
+
+%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{center}

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

 \hline
 \hline

-Embedding &   \multicolumn{2}{|c|}{Adaptive} & Fixed & 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 &   + STC &  + sample & 10\% & 10\%\\ 
+Embedding & Fixed &   \multicolumn{3}{|c|}{Adaptive (about 6.35\%)}  & Fixed & Adapt. & Fixed & Adapt. & Fixed & Adapt. & Fixed & Adapt. \\

 \hline
 \hline

-AUMP & 0.39  & 0.33  & 0.22     &  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 & 0.47 & 0.44 & 0.35     & 0.48     \\
+%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{center}
 \caption{Steganalysing STABYLO\label{table:steganalyse}} 
 \end{center}
 \caption{Steganalysing STABYLO\label{table:steganalyse}} 

-\end{table}
-
-
-Results show that our approach is more easily detectable than HUGO, which
-is the most secure steganographic tool, as far as we know. However due to its 
-huge number of features integration, it is not lightweight.   
-
+\end{table*}
+
+
+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. 
+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 
+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.