2 This section aims at justifying the lightweight attribute of our approach.
3 To be more precise, we compare the complexity of our schemes to some of
4 current state of the art of
5 steganographic schemes, namely HUGO~\cite{DBLP:conf/ih/PevnyFB10},
6 WOW~\cite{conf/wifs/HolubF12}, and UNIWARD~\cite{HFD14}.
7 Each of these schemes starts with the computation of the distortion cost
8 for each pixel switch and is later followed by the STC algorithm.
9 Since this last step is shared by all,
10 we separately evaluate this complexity.
11 In all the remainder of this section, we consider a $n \times n$ square image.
13 First of all, HUGO starts with computing the second order SPAM Features.
14 This steps is in $\theta(n^2 + 2\times 343^2)$ due to the computation
15 of the difference arrays and next of the 686 features (of size 343).
16 Next for each pixel, the distortion measure is calculated by +1/-1 modifying
17 its value and computing again the SPAM
18 features. Pixels are thus selected according to their ability to provide
19 an image whose SPAM features are close to the original ones.
20 The algorithm thus computes a distance between each feature
21 and the original ones,
22 which is at least in $\theta(343)$, and an overall distance between these
23 metrics, which is in $\theta(686)$. Computing the distance is thus in
24 $\theta(2\times 343^2)$ and this modification
25 is thus in $\theta(2\times 343^2 \times n^2)$.
26 Ranking these results may be achieved with a quick sort, which is in
27 $\theta(2 \times n^2 \ln(n))$ for data of size $n^2$.
28 The overall complexity of the pixel selection is finally
29 $\theta(n^2 +2 \times 343^2 + 2\times 343^2 \times n^2 + 2 \times n^2 \ln(n))$, \textit{i.e},
30 $\theta(2 \times n^2(343^2 + \ln(n)))$.
35 Let us focus now on WOW.
36 This scheme starts to compute the residual
37 of the cover as a convolution product which is in $\theta(n^2\ln(n^2))$.
38 The embedding suitability $\eta_{ij}$ is then computed for each pixel
39 $1 \le i,j \le n$ thanks to a convolution product again.
40 We thus have a complexity of $\theta(n^2 \times n^2\ln(n^2))$.
41 Moreover the suitability is computed for each wavelet level
43 This distortion computation step is thus in $\theta(6n^4\ln(n))$.
44 Finally a norm of these three values is computed for each pixel
45 which adds to this complexity the complexity of $\theta(n^2)$.
46 To summarize, the complixity is in $\theta(6n^4\ln(n) +n^2)$
48 What follows details the complexity of the distortion evaluation of the
49 UNIWARD scheme. This one is based to a convolution product $W$ of two elements
50 of size $n$ and is again in $\theta(n^2 \times n^2\ln(n^2))$,
52 these $W$ which is in $\theta(n^2)$.
53 This distortion computation step is thus in $\theta(6n^4\ln(n) + n^2)$.
56 Our edge selection is based on a Canny filter. When applied on a
57 $n \times n$ square image, the noise reduction step is in $\theta(5^3 n^2)$.
58 Next, let $T$ be the size of the Canny mask.
59 Computing gradients is in $\theta(4Tn^2)$ since derivatives of each direction (vertical or horizontal)
60 are in $\theta(2Tn^2)$.
61 Finally, thresholding with hysteresis is in $\theta(n^2)$.
62 The overall complexity is thus in $\theta((5^3+4T+1)n^2)$.
68 We are then left to express the complexity of the STC algorithm.
69 According to~\cite{DBLP:journals/tifs/FillerJF11}, it is
70 in $\theta(2^h.n)$ where $h$ is the size of the duplicated
71 matrix. Its complexity is thus negligible compared with the embedding map
74 The Fig.~\ref{fig:compared}
75 summarizes the complexity of the embedding map construction, for
76 WOW/UNIWARD, HUGO, and STABYLO. It deals with square images
77 of size $n \times n$ when $n$ ranges from
78 512 to 4096. The $y$-coordinate is expressed in a logarithm scale.
79 It shows that the complexity of all the algorithms
80 is dramatically larger than the one of the STABYLO scheme.
81 Thanks to these complexity results, we claim that our approach is lightweight.
84 \includegraphics[scale=0.4]{complexity}
86 \caption{Complexity evaluation of WOW/UNIWARD, HUGO, and STABYLO.}