FAQ
... on the paper "Color
Image Steganalysis Based on Steerable Gaussian Filters Bank,"
Hasan
Abdulrahman, Marc Chaumont, Philippe Montesinos, and Baptiste Magnier,
"Color Image Steganalysis Based on Steerable Gaussian Filters Bank," IH&MMSec'2016,
in Proceedings of the 4th ACM workshop on Information Hiding and
Multimedia Security, Vigo, Galicia, Spain, 6 pages, June 2022, 2016.
Acceptance Rate = 36.2%. pdf. slides. Matlab
code.
This page give additional information about the paper and the
technical part related to what we proposed.
About the dimension of the feature vector:
The first set, produced by [10], is made of
18,157 features.
The second set is made of 4406 features and comes from SPAM features
from
 the gradient magnitude image of each channel (R, G, B) ; it gives
2808
features;
 the derivative image (related to the tangent vector) for each channel
(R,
G, B) ; it gives 1598 features.
More details about the
second set:
 The cooccurences matrices are computed with triplets which means
that
cooccurences have (2xT+1)^3 bins.
 4 matrices are computed (horizontal right, and left, and vertical
right,
and left) and then summed.
 4 others matrices are computed (horizontal right, and left, and
vertical
right and left) and then summed.
 The two matrices are then concatenated in a feature vector.
 For each T values (except for the derivative image with T=3), the
feature
vectors of each channel are
concatenated.
Thus:
 For the gradient magnitude, T is equal to 2 or 3. This leads to a
dimension = 3x(2x5^3) + 3x(2x7^3) = 3x250 + 3x686 = 2808 features,
 For the derivative image, T is equal to 1, 2 or 3. This leads to a
dimension = 3x(2x3^3) + 3x(2x5^3) + (2x7^3) = 3x54+3x250+686 = 1598
features.
Note that
for T=3, we sum the 3
cooccurrences matrices (from R, G, B channels) instead of concatenate
them
(indeed, otherwise, bins values are too small for T=3).
