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