abstract, conclusion

[HindawiJournalOfChaos.git] / IH / IHcomputerJournal.tex

\section{Steganography: a class of secure and robust algorithms~\cite{bcg11:ij}}

This section is devoted to the recall of our contribution published in The
Computer Journal~\cite{bcg11:ij}. 
This article is a review of the researches presented
in the previous sections. Additionnaly, it investigates robustness aspects of the 
process:
applications in frequency domains (namely DWT and DCT embedding) \r
are formalized and corresponding experiments are \r
given~\cite{bcg11:ij}. \r
Such a study shows the applicability of the whole approach.\r
\r
\r
\r
\r
\subsection{Discovering another relevant mode}\r
\label{sec:applications}\r
%\input{applications}\r

We can conclude from the previously summarized article
that we are left to provide:
\begin{itemize}
\item an u.d. strategy-adapter that is independent\r
from the cover, 
\item an image mode $f_l$ whose iteration\r
graph $\Gamma(f_l)$ is strongly connected and whose Markov\r
matrix is doubly stochastic.\r
\end{itemize}
We have recalled in the previous section that the $\textit{CIIS}(K,y,\alpha,l)$ strategy adapter \r
has the required properties.\r
In all the experiments provided in~\cite{bcg11:ij}, parameters $K$ and $\alpha$ are randomly \r
chosen in $] 0, 1[$ and $] 0, 0.5[$ respectively, while the 
number of iteration is set to $4\times lm$, where $lm$ is the number of LSCs \r
that depends on the domain.  \r
 \r
\cite{bcg11:ij} has then used the iterative approach of Section~\ref{sec:prac} to generate image\r
modes $f_l$ such that $\Gamma(f_l)$ is strongly connected, which has
been proposed in~\cite{bcgr11:ip} and recalled in the first part
of this manuscript. Among these\r
maps, it is obvious to check which verifies or not the doubly\r
stochastic constrain.\r
We have already stated that the negation mode matches these hypotheses, so it is relevant in that context.
As a second example, we have considered in~\cite{bcg11:ij} the mode\r
$f_l: \mathds{B}^l \rightarrow \mathds{B}^l$ s.t. its $i$-th component is\r
defined by\r
\begin{equation}\r
\label{eq:fqq}\r
{f_l}(x)_i =\r
\left\{\r
\begin{array}{l}\r
\overline{x_i} \textrm{ if $i$ is odd} \\\r
x_i \oplus x_{i-1} \textrm{ if $i$ is even.}\r
\end{array}\r
\right.\r
\end{equation}\r
\r
Thanks to Theorem~\ref{th:stego}, we have deduced in~\cite{bcg11:ij} that its iteration graph \r
$\Gamma(f_l)$ is strongly connected, and have finally proven\r
that its Markov chain is doubly stochastic by induction on the length $l$.\r
\r
\r
\subsection{dhCI in frequency domains}
We recall in this section the experimental protocol applied in~\cite{bcg11:ij}.

\subsubsection{DWT embedding}\r
%\input{dwt}\r
\r
We have firstly explained in~\cite{bcg11:ij}  how the dhCI dissimulation can be applied in\r
the discrete wavelets transform domain (DWT).\r
The Daubechies family of wavelets has been chosen: \r
% a bitmap file of the famous Lena is converted into its Daubechies-1 DWT\r
% coefficients, which are altered by chaotic iterations. \r
each DWT decomposition depends on a decomposition level and a coefficient\r
matrix (Figure~\ref{fig:DWTs}): $\textit{LL}$ means approximation coefficient,\r
when $\textit{HH},\textit{LH},\textit{HL}$ denote respectively diagonal, \r
vertical, and horizontal detail coefficients. \r
For example, the DWT coefficient \textit{HH}2 is the matrix equal to the \r
diagonal detail coefficient of the second level of decomposition of the image.\r
\r
\begin{figure}[h!]\r
\centerline{\r
\includegraphics[width=4cm]{IH/CompJ/DWTs.eps}\r
}\r
\caption{Wavelets coefficients.}\r
\label{fig:DWTs}\r
\end{figure}\r
\r
\r
\r
\r
\r
The choice of the detail level is motivated by finding\r
a good compromise between robustness and invisibility.\r
Choosing low or high frequencies in DWT domain leads either to a very\r
fragile watermarking without robustness (especially when facing a\r
JPEG2000 compression attack) or to a large degradation of the host\r
content. \r
In order to have a robust but discrete DWT embedding, \r
the second detail level \r
(\textit{i.e.}, $\textit{LH}2,\textit{HL}2,\textit{HH}2$) \r
that corresponds to the middle frequencies,\r
has been retained in~\cite{bcg11:ij}.\r
\r
\r

\r
Let us consider the Daubechies wavelet coefficients of a third\r
level decomposition as represented in Figure~\ref{fig:DWTs}. \r
We then have translated these float coefficients into their 32-bits values, and have
defines in~\cite{bcg11:ij} the  significance function $u$ that associates to any index $k$ in this sequence of bits the following numbers:\r
\begin{itemize}\r
\item $u^k = -1$ if $k$ is one of the three last bits of any index of\r
  coefficients in  $\textit{LH}2$, $\textit{HL}2$, or in $\textit{HH}2$;\r
\item $u^k = 0$ if $k$ is an index of a coefficient in  \r
  $\textit{LH}1$, $\textit{HL}1$, or in $\textit{HH}1$;\r
\item $u^k = 1$ otherwise.\r
\end{itemize}\r
\r
According to the definition of significance of coefficients \r
(Def.~\ref{def:msc,lsc}), if $(m,M)$ is $(-0.5,0.5)$,  LSCs are the\r
last three bits of coefficients in \r
$\textit{HL}2$, $\textit{HH}2$, and $\textit{LH}2$.\r
Thus, decomposition and recomposition functions are fully defined\r
and dhCI dissimulation scheme can now be applied.\r
\r
Figure \ref{fig:DWT} shows the result of a\r
dhCI dissimulation embedding into DWT domain. \r
The original is the image 5007 of the BOSS contest~\cite{Boss10}.\r
Watermark $y$ is given in Fig.~\ref{(b) Watermark}.\r
From a random selection of 50 images into the database from the BOSS \r
contest~\cite{Boss10}, we have applied in~\cite{bcg11:ij} the dhCI algorithm 
with mode $f_l$ \r
defined in the previous section and with the negation mode. \r
\r
\begin{figure}[ht]\r
  \centering\r
  \subfigure[Original Image.]{\includegraphics[width=5cm]\r
    {IH/CompJ/5007.eps}\label{(a) Original 5007}}\hspace{1cm}\r
  \subfigure[Watermark $y$.]{\includegraphics[width=1cm]{IH/CompJ/invader.eps}\label{(b) Watermark}}\hspace{1cm}\r
  \subfigure[Watermarked Image.]{\includegraphics[width=5cm]{IH/CompJ/5007_bis.eps}\label{(c) Watermarked 5007}}\r
\r
\caption{Data hiding in DWT domain}\r
\label{fig:DWT}\r
\end{figure}\r
\r
\r
\subsubsection{DCT embedding}\r
%\input{dct}

We have then explored the discrete cosinus transform (DCT) frequency domain embedding in~\cite{bcg11:ij}, by
following the protocol detailed below.
\r
Let us denote by $x$ the original image of size $H \times L$, and by $y$\r
the hidden message, supposed here to be a binary image of size $H' \times L'$. %\r
The image $x$ is transformed from the spatial\r
domain to DCT domain frequency bands,\r
in order to embed $y$ inside it.  \r
To do so, the host image is firstly divided into $8 \times 8$\r
image blocks as given below:\r
$$x = \bigcup_{k=1}^{H/8} \bigcup_{k'=1}^{L/8} x(k,k').$$\r
Thus, for each image block,\r
a DCT is performed and the coefficients in the frequency bands \r
are obtained as follows:\r
$x_{DCT}(m;n) = DCT(x(m;n))$.\r
\r
To define a discrete but robust scheme, only the three following coefficients of each $8 \times 8$ block in position $(m,n)$ has 
been possibly modified in~\cite{bcg11:ij}: $x_{DCT}(m;n)_{(3,1)},$ $x_{DCT}(m;n)_{(2,2)},$ or $x_{DCT}(m;n)_{(1,3)}$. \r
This choice can be reformulated as follows. \r
Coefficients of each DCT matrix are re-indexed by using a southwest/northeast diagonal, such that $i_{DCT}(m,n)_1 = x_{DCT}(m;n)_{(1,1)}$, $i_{DCT}(m,n)_2 = x_{DCT}(m;n)_{(2,1)}$, $i_{DCT}(m,n)_3 = x_{DCT}(m;n)_{(1,2)}$, $i_{DCT}(m,n)_4 = x_{DCT}(m;n)_{(3,1)}$, ..., and $i_{DCT}(m,n)_{64} =$  $ x_{DCT}(m;n)_{(8,8)}$.\r
So the signification function can be defined in this context by:\r
\begin{itemize}\r
\item if $k$ mod $64 \in \{1,2,3\}$ and $k\leqslant H\times L$, then $u^k=1$;\r
\item else if $k$ mod $64 \in \{4, 5, 6\}$ and $k\leqslant H\times L$, then $u^k=-1$;\r
\item else $u^k = 0$.\r
\end{itemize}\r
The significance of coefficients are obtained for instance with \r
$(m,M)=(-0.5,0.5)$ leading to the definitions of MSCs, LSCs, and passive coefficients.\r
Thus, decomposition and recomposition functions are fully defined
 and dhCI dissimulation scheme has then been applied in~\cite{bcg11:ij}.\r
\r
\r
\r
\r
\subsection{Image quality}\r
%\input{quality}\r
This section focuses on measuring visual quality of our steganographic method.\r
Traditionally, this is achieved by quantifying the similarity \r
between the modified image and its reference image.\r
The Mean Squared Error (MSE) and the Peak Signal to Noise\r
Ratio (PSNR) are the most widely known tools that provide such a metric.\r
However, both of them do not take into account Human Visual System (HVS)\r
properties. \r
Recent works~\cite{EAPLBC06,SheikhB06,PSECAL07,MB10} have tackled this problem \r
by creating new metrics. Among them, what follows focuses on PSNR-HVS-M~\cite{PSECAL07} and BIQI~\cite{MB10}, considered as advanced visual quality metrics.  \r
The former efficiently combines PSNR and  visual between-coefficient contrast masking of DCT basis functions based on HVS. This metric has \r
been computed in~\cite{bcg11:ij}
 by using the implementation given at~\cite{psnrhvsm11}.\r
The latter allows to get a blind image quality assessment measure, \r
\textit{i.e.}, without any knowledge of the source distortion.\r
Its implementation is available at~\cite{biqi11}.\r
\r
\r
\begin{table}\r
\begin{center}\r
\begin{tabular}{|c|c|c|c|c|}\r
\hline\r
Embedding & \multicolumn{2}{|c|}{DWT} \r
 & \multicolumn{2}{|c|}{DCT} \\\r
\hline\r
Mode & $f_l$ & neg. & $f_l$ & neg. \\\r
\hline\r
PSNR & 42.74     & 42.76     &  52.68      &  52.41   \\\r
\hline\r
PSNR-HVS-M & 44.28  & 43.97 & 45.30 & 44.93 \\\r
\hline\r
BIQI & 35.35 & 32.78 & 41.59 & 47.47 \\\r
\hline\r
\end{tabular}\r
\end{center}\r
\caption{Quality measures of our steganography approach~\cite{bcg11:ij}}
\label{table:quality} \r
\end{table}\r
\r
\r
\r
Results of the image quality metrics obtained in~\cite{bcg11:ij}\r
are summarized in Table~\ref{table:quality}.\r
In wavelet domain, the PSNR values obtained in~\cite{bcg11:ij} are comparable to other approaches\r
(for instance, PSNR are 44.2 in~\cite{TCL05} and 46.5 in~\cite{DA10}), \r
but  a real improvement for the discrete cosine embeddings is obtained \r
(PSNR is 45.17 for~\cite{CFS08}, it is always lower than 48 for~\cite{Mohanty:2008:IWB:1413862.1413865}, and always lower than 39 for~\cite{MK08}).\r
Among steganography approaches that evaluate PSNR-HVS-M, results of our approach \r
are convincing. Firstly, optimized method developed along~\cite{Randall11} has a PSNR-HVS-M equal to 44.5 whereas our approach, with a similar PSNR-HVS-M, should be easily improved by considering optimized mode. Next, \r
another approach~\cite{Muzzarelli:2010} have higher PSNR-HVS-M, certainly, but\r
this work does not address robustness evaluation whereas the study presented in~\cite{bcg11:ij} is complete.\r
Finally, as far as we know, \cite{bcg11:ij} is the first one that has evaluated the BIQI metric in a 
steganographic context. \r
\r
 \r
\r
With all this material, we have then evaluated the robustness of our \r
approach in~\cite{bcg11:ij}. \r
\r
\r
\r
\subsection{Robustness}\r
%\input{robustness}\r
Previous sections have formalized frequential domains embeddings and\r
has focused on the negation and $f_l$ modes.\r
In the robustness given in this continuation, {dwt}(neg), \r
{dwt}(fl), {dct}(neg), and {dct}(fl) \r
respectively stand for the DWT and DCT embedding \r
with the negation mode and with this instantiated mode.\r
 \r
For each experiment presented in~\cite{bcg11:ij}, a set of 50 images is randomly extracted \r
from the database taken from the BOSS contest~\cite{Boss10}. \r
Each cover is a $512\times 512$ grayscale digital image and the watermark $y$ \r
is given in Figure~\ref{(b) Watermark}. \r
Testing the robustness of the approach is achieved in~\cite{bcg11:ij} by successively applying\r
on watermarked images attacks like cropping, compression, and geometric \r
transformations.\r
Differences between \r
$\hat{y}$ and $\varphi_m(z)$ have then been\r
computed. Behind a given threshold rate, the image is said to be watermarked.  \r
Finally, discussion on metric quality of the approach given in~\cite{bcg11:ij} is recalled in \r
Section~\ref{sub:roc}.\r
\r
 \r
\r
\subsubsection{Robustness against cropping}
\r
Robustness of the approach is  evaluated by\r
applying different percentage of cropping: from 1\% to 81\%.\r
Results obtained in~\cite{bcg11:ij} are recalled in Figure~\ref{Fig:atck:dec}. 
Figure~\ref{Fig:atq:dec:img}\r
gives the cropped image \r
where 36\% of the image is removed, while Figure~\ref{Fig:atq:dec:curves} 
presents effects of such an attack.\r
From this experiment, we have concluded in~\cite{bcg11:ij} that all embeddings have similar \r
behaviors.\r
All the percentage differences are so far less than 50\% \r
(which is the mean random error) and thus robustness is established.\r
\r
\r
\r
\begin{figure}[ht]\r
  \centering\r
  \subfigure[Cropped image]{\includegraphics[width=0.24\textwidth]\r
    {IH/CompJ/5007_dec_307.eps}\label{Fig:atq:dec:img}}\hspace{2cm}\r
  \subfigure[Cropping effects]{\r
\includegraphics[width=0.5\textwidth]{IH/CompJ/atq-dec.eps}\label{Fig:atq:dec:curves}}\r
\caption{Cropping results}\r
\label{Fig:atck:dec}\r
\end{figure}\r
\r
\r
\subsubsection{Robustness against compression}\r
\r
Robustness against compression is addressed\r
by studying both JPEG  and JPEG 2000 image compressions.\r
Results  obtained in~\cite{bcg11:ij}  are respectively presented in Fig.~\ref{Fig:atq:jpg:curves}\r
and Fig.~\ref{Fig:atq:jp2:curves}.\r
Without surprise, DCT embedding which is based on DCT \r
(as JPEG compression algorithm is) is  more \r
adapted to JPEG compression than DWT embedding.\r
Furthermore, we have a similar behavior for the JPEG 2000 compression algorithm, which is based on wavelet encoding: DWT embedding naturally outperforms\r
DCT one in that case.  \r
\r
\r
\begin{figure}[ht]\r
  \centering\r
  \subfigure[JPEG effects]{\r
\includegraphics[width=0.48\textwidth]{IH/CompJ/atq-jpg.eps}\label{Fig:atq:jpg:curves}}\r
  \subfigure[JPEG 2000 effects]{\r
\includegraphics[width=0.48\textwidth]{IH/CompJ/atq-jp2.eps}\label{Fig:atq:jp2:curves}}\r
\caption{Compression results}\r
\label{Fig:atck:comp}\r
\end{figure}\r
\r
\r
\r
\subsubsection{Robustness against contrast and sharpness}\r

Contrast and Sharpness adjustments belong to the classical set of \r
filtering image attacks.\r
Results of such attacks are presented in \r
Fig.~\ref{Fig:atq:fil} where \r
Fig.~\ref{Fig:atq:cont:curve} and Fig.~\ref{Fig:atq:sh:curve} summarize \r
effects of contrast and sharpness adjustment respectively~\cite{bcg11:ij}. \r
\r
\r
\begin{figure}[ht]\r
  \centering\r
  \subfigure[Contrast effects]{\r
\includegraphics[width=0.48\textwidth]{IH/CompJ/atq-contrast.eps}\label{Fig:atq:cont:curve}}\r
  \subfigure[Sharpness effects]{\r
\includegraphics[width=0.48\textwidth]{IH/CompJ/atq-flou.eps}\label{Fig:atq:sh:curve}}\r
\caption{Filtering results}\r
\label{Fig:atq:fil}\r
\end{figure}\r
\r
\subsubsection{Robustness against geometric transformations}\r

Among geometric transformations, we have focused in~\cite{bcg11:ij} on  \r
rotations, \textit{i.e.}, when two opposite rotations \r
of angle $\theta$ are successively applied around the center of the image.\r
In these geometric transformations,  angles range from 2 to 20 \r
degrees.  \r
Results obtained in~\cite{bcg11:ij} are summed up in Figure~\ref{Fig:atq:rot}: Fig.~\ref{Fig:atq:rot:img}\r
gives the image of a rotation of 20 degrees whereas\r
Fig.~\ref{Fig:atq:rot:curve} presents effects of such an attack.\r
It is not a surprise that results are better for DCT embeddings: this approach\r
is based on cosine as rotation is. \r
\r
\r
\r
\begin{figure}[ht]\r
  \centering\r
  \subfigure[20 degrees rotation image]{\r
    \includegraphics[width=0.25\textwidth]{IH/CompJ/5007_rot_10.eps}\label{Fig:atq:rot:img}}\r
  \subfigure[Rotation effects]{\r
\includegraphics[width=0.5\textwidth]{IH/CompJ/atq-rot.eps}\label{Fig:atq:rot:curve}}\r
\r
\caption{Rotation attack results}\r
\label{Fig:atq:rot}\r
\end{figure}\r
\r
\subsection{Evaluation of the  Embeddings}
\label{sub:roc}\r

We are then left to set a convenient threshold that is accurate to \r
determine whether an image is watermarked or not.\r
Starting from a set of 100 images selected among the Boss image panel,\r
we have computed in~\cite{bcg11:ij} the following three sets: \r
the one with all the watermarked images $W$,\r
the one with all successively watermarked and attacked images $\textit{WA}$,\r
and the one with only the attacked images $A$.\r
Notice that the 100 attacks for each image\r
are selected among these detailed previously.\r
\r
\r
For each threshold $t \in \llbracket 0,55 \rrbracket$ and a given image \r
$x \in \textit{WA} \cup A$, \r
differences on DCT have been computed in~\cite{bcg11:ij}. 
The image has been claimed as watermarked\r
if these differences are less than the threshold. \r
\begin{itemize}
\item In the positive case and if $x$ really belongs to \r
$\textit{WA}$ it is a True Positive (TP) case.  \r
\item In the negative case but if $x$ belongs to \r
$\textit{WA}$, it is a False Negative (FN) case.  \r
\item In the positive case but if $x$  belongs to \r
$\textit{A}$, it is a False Positive (FP) case.  \r
\item Finally, in the negative case and if $x$ belongs to\r
$\textit{A}$, it is a True Negative (TN).  \r
\end{itemize}
The True (resp. False) Positive Rate  (TPR) (resp. FPR) has thus been computed \r
by dividing the number of TP (resp. FP) by 100.\r
\r
\begin{figure}[ht]\r
\begin{center}\r
\includegraphics[width=9cm]{IH/CompJ/ROC.eps}\r
\end{center}\r
\caption{ROC curves for DWT or DCT embeddings}\label{fig:roc:dwt}\r
\end{figure}\r
\r
Figure~\ref{fig:roc:dwt} recalled the obtained Receiver Operating Characteristic (ROC) \r
curve. \r
For the DWT, it shows that best results are obtained when the threshold \r
is 45\% for the dedicated function (corresponding to the point (0.01, 0.88))\r
and  46\% for the negation function (corresponding to (0.04, 0.85)).\r
It allows to conclude that each time LSCs differences between\r
a watermarked image and another given image  $i'$ are less than 45\%, we can claim that \r
$i'$ is an attacked version of the original watermarked content.\r
For the two DCT embeddings, best results have been obtained when the threshold \r
is 44\% (corresponding to the points (0.05, 0.18) and (0.05, 0.28)).\r
\r
We have thus conclude some confidence intervals for all the evaluated 
attacks in~\cite{bcg11:ij}. The\r
approach is resistant to:\r
\begin{itemize}\r
\item all the croppings where percentage is less than 85;\r
\item compressions where quality ratio is greater \r
  than 82 with DWT embedding and \r
  where quality ratio is greater than 67 with DCT one;\r
\item contrast when strengthening belongs to $[0.76,1.2]$ \r
(resp. $[0.96,1.05]$)  in DWT (resp. in DCT) embedding;\r
\item all the rotation attacks with DCT embedding and a rotation where\r
angle is less than 13 degrees with DWT one.\r
\end{itemize}\r
\r
\r
% \subsection{Conclusion}\label{sec:concl}\r
% %\input{conclusion}\r
% This paper has proposed a new class of secure and robust information hiding \r
% algorithms.\r
% It has been entirely formalized, thus allowing both its theoretical security \r
% analysis, and the computation of numerous variants encompassing spatial and \r
% frequency domain embedding.\r
% After having presented the general algorithm with detail, we have given\r
% conditions for choosing mode and strategy-adapter making the whole\r
% class  stego-secure or $\epsilon$-stego-secure.\r
% To our knowledge, this is the first time such a result has been established.\r
% 
% \r
% Applications in frequency domains (namely DWT and DCT domains) have finally be\r
% formalized.\r
% Complete experiments have allowed us \r
% first to evaluate how invisible is the steganographic method (thanks to the PSNR computation) and next to verify the robustness property against attacks.\r
% Furthermore, the use of ROC curves for DWT embedding have revealed very high rates\r
% between True positive and False positive results.\r
% \r
% In future work, our intention is to find the best image mode with respect to  \r
% the combination between  DCT and DWT based steganography\r
% algorithm. Such a combination topic has already been addressed\r
% (\textit{e.g.}, in~\cite{al2007combined}), but never with objectives\r
% we have set.\r
\r
\r
% Additionally, we will try to discover new topological properties for the dhCI\r
% dissimulation schemes.\r
% Consequences of these chaos properties will be drawn in the context of \r
% information hiding security.\r
% We will especially focus on the links between topological properties and classes\r
% of attacks, such as KOA, KMA, EOA, or CMA.\r
% \r
% Moreover, these algorithms will be compared to other existing ones, among other\r
% things by testing whether these algorithms are chaotic or not.\r
% Finally we plan to verify the robustness of our approach \r
% against statistical steganalysis methods~\cite{GFH06,ChenS08,DongT08,FridrichKHG11a}.\r
% \r%