1 For the whole experiment, a set of 500 images is randomly extracted
2 from the database taken from the BOSS contest~\cite{Boss10}.
3 In this set, each cover is a $512\times 512$
4 grayscale digital image.
7 \subsection{Adaptive Embedding Rate}
9 Two strategies have been developed in our scheme with respect to the rate of
10 embedding which is either \emph{adaptive} or \emph{fixed}.
12 In the former the embedding rate depends on the number of edge pixels.
13 The higher it is, the larger is the message length that can be considered.
14 Practically, a set of edge pixels is computed according to the
15 Canny algorithm with high threshold.
16 The message length is thus defined to be the half of this set cardinality.
17 In this strategy, two methods are thus applied to extract bits that
18 are modified. The first one is a direct application of the STC algorithm.
19 This method is further referred as \emph{adaptive+STC}.
20 The second one randomly choose the subset of pixels to modify by
21 applying the BBS PRNG again. This method is denoted \emph{adaptive+sample}.
22 Notice that the rate between
23 available bits and bit message length is always equal to two.
24 This constraint is indeed induced by the fact that the efficiency
25 of the STC algorithm is unsatisfactory under that threshold.
27 On our experiments and with the adaptive scheme,
28 the average size of the message that can be embedded is 16445.
29 Its corresponds to an average payload of 6.35\%.
34 In the latter, the embedding rate is defined as a percentage between the
35 number of the modified pixels and the length of the bit message.
36 This is the classical approach adopted in steganography.
37 Practically, the Canny algorithm generates a
38 a set of edge pixels with threshold that is decreasing until its cardinality
39 is sufficient. If the set cardinality is more than twice larger than the
40 bit message length an STC step is again applied.
41 Otherwise, pixels are again randomly chosen with BBS.
45 \subsection{Image Quality}
46 The visual quality of the STABYLO scheme is evaluated in this section.
47 Four metrics are computed in these experiments:
48 the Peak Signal to Noise Ratio (PSNR),
49 the PSNR-HVS-M family~\cite{PSECAL07,psnrhvsm11} ,
50 the BIQI~\cite{MB10,biqi11} and
51 the weighted PSNR (wPSNR)~\cite{DBLP:conf/ih/PereiraVMMP01}.
52 The first one is widely used but does not take into
53 account Human Visual System (HVS).
54 The other last ones have been designed to tackle this problem.
58 \begin{tabular}{|c|c|c||c|c|}
60 Schemes & \multicolumn{3}{|c|}{STABYLO} & HUGO\\
62 Embedding & \multicolumn{2}{|c||}{Adaptive} & Fixed & Fixed \\
64 Rate & + STC & + sample & 10\% & 10\%\\
66 PSNR & 66.55 & 63.48 & 61.86 & 64.65 \\
68 PSNR-HVS-M & 78.6 & 75.39 & 72.9 & 76.67\\
70 BIQI & 28.3 & 28.28 & 28.4 & 28.28\\
72 wPSNR & 86.43& 80.59 & 77.47& 83.03\\
76 \caption{Quality Measures of Steganography Approaches\label{table:quality}}
79 Let us give an interpretation of these experiments.
80 First of all, the adaptive strategy produces images with lower distortion
81 than the one of images resulting from the 10\% fixed strategy.
82 Numerical results are indeed always greater for the former strategy than
83 for the latter, except for the BIQI metrics where differences are not relevant.
84 These results are not surprising since the adaptive strategy aims at
85 embedding messages whose length is decided according to a higher threshold
86 into the edge detection.
87 Let us focus on the quality of HUGO images: with a given fixed
89 HUGO always produces images whose quality is higher than the STABYLO's one.
90 However, our approach nevertheless provides better results with the strategy
91 adaptive+STC in a lightweight manner, as motivated in the introduction.
94 Let us now compare the STABYLO approach with other edge based steganography
95 schemes with respect to the image quality.
96 First of all, the Edge Adaptive
97 scheme detailed in~\cite{Luo:2010:EAI:1824719.1824720}
98 executed with a 10\% embedding rate
99 has the same PSNR but a lower wPSNR than our:
100 these two metrics are respectively equal to 61.9 and 68.9.
101 Next both the approaches~\cite{DBLP:journals/eswa/ChenCL10,Chang20101286}
102 focus on increasing the payload while the PSNR is acceptable, but do not
103 give quality metrics for fixed embedding rate from a large base of images.
104 Our approach outperforms the former thanks to the introduction of the STC
110 \subsection{Steganalysis}
114 The quality of our approach has been evaluated through the two
115 AUMP~\cite{Fillatre:2012:ASL:2333143.2333587}
116 and Ensemble Classifier~\cite{DBLP:journals/tifs/KodovskyFH12} based steganalysers.
117 Both aims at detecting hidden bits in grayscale natural images and are
118 considered as the state of the art of steganalysers in spatial domain~\cite{FK12}.
119 The former approach is based on a simplified parametric model of natural images.
120 Parameters are firstly estimated and a adaptive Asymptotically Uniformly Most Powerful
121 (AUMP) test is designed (theoretically and practically) to check whether
122 a natural image has stego content or not.
123 In the latter, the authors show that the
124 machine learning step, (which is often
125 implemented as support vector machine)
126 can be a favorably executed thanks to an Ensemble Classifiers.
132 \begin{tabular}{|c|c|c|c|c|}
134 Schemes & \multicolumn{3}{|c|}{STABYLO} & HUGO\\
136 Embedding & \multicolumn{2}{|c|}{Adaptive} & Fixed & Fixed \\
138 Rate & + STC & + sample & 10\% & 10\%\\
140 AUMP & 0.39 & 0.33 & 0.22 & 0.50 \\
142 Ensemble Classifier & 0.47 & 0.44 & 0.35 & 0.48 \\
147 \caption{Steganalysing STABYLO\label{table:steganalyse}}
151 Results show that our approach is more easily detectable than HUGO, which
152 is the most secure steganographic tool, as far as we know. However due to its
153 huge number of features integration, it is not lightweight.