Steganalysis of content-adaptive JPEG steganography
based on Gauss partial derivative filter bank
Yi Zhang,
a,b
Fenlin Liu,
a,b
Chunfang Yang,
a,b,
* Xiangyang Luo,
a,b
Xiaofeng Song,
c
and Jicang Lu
b
a
State Key Laboratory of Mathematical Engineering and Advanced Computing, Zhengzhou, Henan, China
b
Zhengzhou Science and Technology Institute, Zhengzhou, Henan, China
c
Xi’an Communication Institute, Xi’an, Shanxi, China
Abstract. A steganalysis feature extraction method based on Gauss partial derivative filter bank is proposed in
this paper to improve the detection performance for content-adaptive JPEG steganography. Considering that the
embedding changes of content-adaptive steganographic schemes are performed in the texture and edge
regions, the proposed method generates filtered images comprising rich texture and edge information using
Gauss partial derivative filter bank, and histograms of absolute values of filtered subimages are extracted
as steganalysis features. Gauss partial derivative filter bank can represent texture and edge information in multi-
ple orientations with less computation load than conventional methods and prevent redundancy in different fil-
tered images. These two properties are beneficial in the extraction of low-complexity sensitive features. The
results of experiments conducted on three selected modern JPEG steganographic schemes—uniform embed-
ding distortion, JPEG universal wavelet relative distortion, and side-informed UNIWARD —indicate that the pro-
posed feature set is superior to the prior art feature sets—discrete cosine transform residual, phase aware rich
model, and Gabor filter residual.
© 2017 SPIE and IS&T [DOI: 10.1117/1.JEI.26.1.013011]
Keywords: content-adaptive JPEG steganography; steganalysis feature; Gauss partial derivative filter bank; histograms.
Paper 160968 received Nov. 16, 2016; accepted for publication Jan. 12, 2017; published online Feb. 8, 2017.
1 Introduction
Content-adaptive steganography is the most successful form
of modern steganographic schemes for digital images.
1
In
these schemes, a cost is first assigned to the changing of
each cover element [e.g., pixel or discrete cosine transform
(DCT) coefficient] according to the distortion functions, and
then the total distortion is minimized as the sum of costs
of all modified elements with syndrome-trellis codes.
2
Elements in the texture and edge regions that are difficult
to describe using statistical models are usually assigned
lower costs, which makes the embedding changes more dif-
ficult to be detected.
3
Many schemes of this form have been
proposed based on various distortion functions (the major
differences between these schemes), such as highly undetect-
able steGO,
4,5
wavelet obtained weights,
6
spatial
UNIWARD,
7
and high pass, low pass and low pass
8
in spatial
domain and normalized perturbed quantization,
9
uniform
embedding distortion (UED),
10
JPEG universal wavelet
relative distortion (JUNIWARD),
7
and side-informed
UNIWARD (SIUNIWARD)
7
in JPEG domain.
In recent years, content-adaptive JPEG steganographic
schemes have been best detected with features assembled
as histograms of filtered images, which are generated by con-
volving the decompressed JPEG image with filters. The
state-of-the-art feature sets are discrete cosine transform
residual (DCTR),
11
phase aware rich model (PHARM),
12
and Gabor filter residual (GFR).
13
The design of the filters
is the major difference between them. In the DCTR feature
set, 64 DCT bases are employed as the filter bank, and 8000-
dimensional features are extracted for detect ion, with the
most appealing aspect being low computational complexity.
In the PHARM feature set, 900 random projections of nine-
pixel predictors are employed as its filter bank, and 12600-
dimensional features are extracted for detection. Although it
has obvious advantages in terms of detection performance
when compared with DCTR, its extraction time is expensive.
In the GFR feature set, 256 Gabor filters with four scales,
two phases, and 32 orientations are employed as the filter
bank, and 17,000-dimensional features are extracted for
detection. Its detection performance is superior to that of
DCTR and PHARM, and its extraction time is between
their extraction times. Moreover, by accumulating in the his-
tograms a quanti ty that bounds the expected absolute distor-
tion of the filtered images, the detection performance of the
above feature sets is significantly improved in Ref. 14.
Because the embedding changes of content-adaptive steg-
anographic schemes are perfor med in the texture and edge
regions, full acquisition of texture and edge information
facilitates extraction of sensitive features. Most conventional
methods capture texture and edge information as completely
as possible based on filters, whose ability to capture such
information directly affects the detection performance of
these methods. As with the Gabor filter in GFR, steerable
filter
15
is also a commonly used and good performance tool
to capture image texture and edge information. One remark-
able characteristic is that it can be constructed in any orien-
tation by taking a suitable linear combination of a small
number of basis filters.
16
This characteristic makes it easy
to capture the texture and edge information in multiple ori-
entations. However, the linear relationship with the basis fil-
ters results in the filtered images generated by steerable
filters with different orientations having redundancies,
which is not conducive to the construction of sensitive
*Address all correspondence to: Chunfang Yang, E-mail: chunfangyang@126
.com
1017-9909/2017/$25.00 © 2017 SPIE and IS&T
Journal of Electronic Imaging 013011-1 Jan∕Feb 2017
•
Vol. 26(1)
Journal of Electronic Imaging 26(1), 013011 (Jan∕Feb 2017)
Downloaded From: http://electronicimaging.spiedigitallibrary.org/pdfaccess.ashx?url=/data/journals/electim/935790/ on 02/14/2017 Terms of Use: http://spiedigitallibrary.org/ss/termsofuse.aspx