1 This section is devoted to the robustness study of our scheme.
2 This one has to ensure that the watermark withstands against
3 different types of active attacks that modify the watermarked image.
5 For the whole experiment, a set of 100 images is randomly extracted
6 from the database taken from the BOSS contest~\cite{Boss10}.
7 In this set, each cover is a $512\times 512$
8 grayscale digital image.
9 The considered watermark $m$ is given in Fig.~\ref{(b) Watermark}.
10 Testing the robustness of the approach is achieved by successively applying
11 on watermarked images attacks like cropping, compression, geometric
12 transformations,\ldots % (from Sec.~\ref{sub:crop} to Sec.~\ref{sub:lsb}).
19 \includegraphics[width=0.06\textwidth]{exp/invader}
21 \caption{The Watermark}
26 Differences between $m$ and $m'$ are
27 computed. Behind a given threshold rate, the image is said to be watermarked.
29 Discussion on metric quality of the approach is finally given.
33 In the remainder of this article, on following figures, the difference percentage
34 corresponds to the distance between the retrieved and the original watermarks.
37 %\subsection{Robustness against Cropping}\label{sub:crop}
38 %\subsection{Against Cropping Attack}\label{sub:crop}
40 Robustness of the approach is evaluated by
41 applying different percentages of cropping: from 0.25\% to 90\%.
42 Results are given in Fig.~\ref{Fig:atck:dec}, %.
43 %, and Fig.~\ref{Fig:atq:dec:img} gives the cropped image
44 %where 36\% of the image is removed.
45 %Fig.~\ref{Fig:atq:dec:curves}
46 which presents effects of such an attack.
47 All the percentage differences are so far less than 97\%
48 and thus robustness is established.
54 % \subfigure[Cropped Image.]{\includegraphics[width=0.24\textwidth]
55 % {5007_dec_307.png}\label{Fig:atq:dec:img}}\hspace{2cm}
56 % \subfigure[Cropped Image.]{\includegraphics[width=0.24\textwidth]
57 % {5007_dec_307}\label{Fig:atq:dec:img}}\hspace{2cm}
58 % \subfigure[Cropping Effect]{
59 \includegraphics[width=0.35\textwidth]{exp/atq-dec}\label{Fig:atq:dec:curves}%}
60 %\includegraphics[width=0.5\textwidth]{exp/atq-dec.eps}\label{Fig:atq:dec:curves}}
61 \caption{Cropping Results}
66 %\subsection{Robustness against compression}\label{sub:comp}
67 %\subsection{Against Compression Attack}\label{sub:comp}
69 Robustness against compression is addressed
70 by studying both JPEG and JPEG 2000 image compression.
71 Results are respectively presented in Fig.~\ref{Fig:atq:jpg:curves}
72 and Fig.~\ref{Fig:atq:jp2:curves}.
73 It is not hard to see that robustness is well established for
74 JPEG2000 compression: for all the ratio larger than 10\%, the watermark
76 However, our scheme is not robust against JPEG compression for a ratio inferior
79 A potential solution in order to improve this result should be to insert the
80 watermark in least significant coefficient of the image described in frequency
81 domain as for example with discrete cosine transform or with wavelet transform.
82 This study will be described in future works.
83 %\JFC{voir comment amméliorer}
86 \subfigure[JPEG Effect]{
87 \includegraphics[width=0.35\textwidth]{exp/atq-jpg}\label{Fig:atq:jpg:curves}}
88 %\includegraphics[width=0.45\textwidth]{exp/atq-jpg.eps}\label{Fig:atq:jpg:curves}}
89 \subfigure[JPEG 2000 Effect]{
90 \includegraphics[width=0.35\textwidth]{exp/atq-jp2}\label{Fig:atq:jp2:curves}}
91 %\includegraphics[width=0.45\textwidth]{exp/atq-jp2.eps}\label{Fig:atq:jp2:curves}}
92 \caption{Compression Results}
98 %\subsection{Robustness against Contrast and Sharpness
99 %Attack}\label{sub:contrast}
100 %\subsection{Against Contrast and Sharpness Attack}\label{sub:contrast}
102 % Contrast and Sharpness adjustment belong to the classical set of
103 % filtering image attacks.
104 % Results of such attacks are presented in
105 % Fig.~\ref{Fig:atq:fil}, where
106 % Fig.~\ref{Fig:atq:cont:curve} and Fig.~\ref{Fig:atq:sh:curve} summarize
107 % effects of contrast and sharpness adjustment respectively.
108 % Robustness is not established for both these types of attacks.
109 % Future works give issues to tackle this problem.
113 % \subfigure[Contrast Effect]{
114 % \includegraphics[width=0.4\textwidth]{exp/atq-contrast}\label{Fig:atq:cont:curve}}
115 % %\includegraphics[width=0.45\textwidth]{exp/atq-contrast.eps}\label{Fig:atq:cont:curve}}
116 % \subfigure[Sharpness Effect]{
117 % \includegraphics[width=0.4\textwidth]{exp/atq-flou}\label{Fig:atq:sh:curve}}
118 % %\includegraphics[width=0.45\textwidth]{exp/atq-flou.eps}\label{Fig:atq:sh:curve}}
119 % \caption{Filtering Results}
120 % \label{Fig:atq:fil}
124 %nom;../../../../BCG10/Oxford11/experiments/images/1247 donne des résultats
127 %\subsection{Against Geometric Transformation}\label{sub:rotation}
128 Among geometric transformations, we focus on
129 rotations, \textit{i.e.}, when two opposite rotations
130 of angle $\theta$ are successively applied around the center of the image.
131 In these geometric transformations, angles range from 2 to 60
133 Results are presented in Fig.~\ref{Fig:atq:rot}. %: Fig.~\ref{Fig:atq:rot:img}
134 % gives the image of a rotation of 20 degrees whereas
135 % Fig.~\ref{Fig:atq:rot:curve} presents the effects of such an attack.
136 Thanks to an efficient embedding, our scheme is resistant to all
137 that type of attacks.
143 % \subfigure[20 degrees Rotation Image]{
144 % \includegraphics[width=0.2\textwidth]{5007_rot_20}\label{Fig:atq:rot:img}}
145 % % \includegraphics[width=0.25\textwidth]{5007_rot_10.eps}\label{Fig:atq:rot:img}}
147 % \subfigure[Rotation Effect]{
148 \includegraphics[width=0.35\textwidth]{exp/atq-rot}\label{Fig:atq:rot:curve}%}
149 %\includegraphics[width=0.45\textwidth]{exp/atq-rot.eps}\label{Fig:atq:rot:curve}}
151 \caption{Rotation Attack Results}
155 %\subsection{Against LSB Transformation}\label{sub:lsb}
156 The first step of this scheme has defined $x$ as the LSB of the host and
157 is thus based on LSB modifications.
158 This part focuses on two types of attacks modifying these LSB sets (see
159 Fig~\ref{Fig:atq:lsb}). The former consists in setting to zero a subset of this one.
160 Results are expressed in Fig.~\ref{Fig:atq:lsb:curve} and
161 show that the scheme is robust, unless 95\% of the LSB is erased.
162 In this case the image is really damaged.
163 The latter consists in applying again this scheme on the watermarked image
164 but with another message.
165 Results of Fig.~\ref{Fig:atq:ci2:curve} show that this scheme is robust
166 against that type of attack, provided the number of iterations is
167 lesser than $1.75$ times the number of pixels.
168 With more iterations, the image is dramatically modified: more than 50\%
169 of the LSB is switched. However, future works present ideas to tackle
175 \subfigure[LSB Erasing Effect]{
176 \includegraphics[width=0.35\textwidth]{exp/atq-lsb}\label{Fig:atq:lsb:curve}}
178 \subfigure[Applying Algorithm twice]{
179 \includegraphics[width=0.35\textwidth]{exp/atq-ci2}\label{Fig:atq:ci2:curve}}
180 %\includegraphics[width=0.45\textwidth]{exp/atq-rot.eps}\label{Fig:atq:rot:curve}}
182 \caption{LSB Modifications}
189 \subsection{Evaluation of the Embeddings}\label{sub:roc}
190 A Receiver Operating Characteristic approach has been implemented to find
191 the most adapted threshold w.r.t. the separation between
192 watermarked images and other ones.
196 \includegraphics[width=0.35\textwidth]{exp/ROC}
197 %\includegraphics[width=7cm]{ROC.eps}
199 \caption{ROC Curves for DWT or DCT Embeddings}\label{fig:roc:dwt}
202 The Figure~\ref{fig:roc:dwt} is the Receiver Operating Characteristic (ROC)
203 curve. This curve is close to the ideal one that is without
204 False Positive and False Negative answer.
205 The threshold with best results is a distance equal to 0,97.
206 With such a value, we can give some confidence intervals
208 evaluated attacks. The
209 approach is resistant to all the cropping where percentage is less than 90,
210 to JPEG200 compression where quality ratio is greater than 5\%,
211 to all the rotation attacks,
212 to LSB erasing when less than 95\% are set to 0,
213 a second application of the scheme with less than 1.75 iterations