Cover Selection Steganography Method Based on Similarity of Image Blocks

Abstract

An advantage of steganography, as opposed to other information hiding techniques, is that the embedder can select a cover image that results in the least detectable stego image. In a previously proposed method, a technique based on block texture similarity was introduced where blocks of cover image were replaced with the similar secret image blocks; then indices of secret image blocks were stored in cover image. In this method, the blocks of secret image are compared with blocks of a set of cover images and the image with most similar blocks to those of the secret image is selected as the best candidate to carry the secret image. Using appropriate features for comparing image blocks, guaranties higher quality of stego images and consequently, allows for higher embedding capacity, less delectability and, enhanced security. Based on this idea, in this paper, an adaptive cover selection steganography method is proposed, that uses statistical features of image blocks and their neighborhood. Using the block neighborhood information, we prevent appearing virtual edges in the sides and corners of the replaced blocks. Our method is examined with feature based and wavelet based steganalysis algorithms. The results prove the effectiveness and benefits of the proposed method.

Full text

Cover Selection Steganography Method Based on
Similarity of Image Blocks
Hedieh Sajedi, Mansour Jamzad
Computer Engineering Department, Sharif University of Technology
A_sajedi@ce.sharif.edu, Jamzad@sharif.edu
Abstract
An advantage of steganography, as opposed to
other information hiding techniques, is that the
embedder can select a cover image that results in the
least detectable stego image. In a previously proposed
method, a technique based on block texture similarity
was introduced where blocks of cover image were
replaced with the similar secret image blocks; then
indices of secret image blocks were stored in cover
image. In this method, the blocks of secret image are
compared with blocks of a set of cover images and the
image with most similar blocks to those of the secret
image is selected as the best candidate to carry the
secret image. Using appropriate features for
comparing image blocks, guaranties higher quality of
stego images and consequently, allows for higher
embedding capacity, less delectability and, enhanced
security. Based on this idea, in this paper, an adaptive
cover selection steganography method is proposed,
that uses statistical features of image blocks and their
neighborhood. Using the block neighborhood
information, we prevent appearing virtual edges in the
sides and corners of the replaced blocks. Our method
is examined with feature based and wavelet based
steganalysis algorithms. The results prove the
effectiveness and benefits of the proposed method.
1. Introduction
Steganography is the art and science of writing
hidden messages in such a way that no one, except the
intended recipient, knows of the existence of the
message. Consider that a transmitter consists of a host
image H, and the message M that a sender hopes to
communicate confidentially. The message can be text,
images, or anything that can be represented by a bit
stream. The host image H is used to embed the message
by using a stego-encoder controlled by a key K. The
key is a shared secret with the intended recipient whose
knowledge of the key enables him to decode the
message from the stego image. The decoding param-
eters are known to both sender and receiver as a shared
secret. The resulting stego image, SI = f (H,M,K), is
transmitted over a channel to the receiver where it is
processed by the stego-decoder using the same key K.
Successful steganography depends upon the carrier
medium not to attract attention. When presence of
stego-content is suspected, the main goal of
steganography is defeated [1]. There is a tradeoff
between the invisibility (imperceptible to naked eye)
and the amount of information that can be hidden in a
given cover image [2]. Steganographic security is
mostly influenced by the type of cover media; the
method for selection of places within the cover that
might be modified; the type of embedding operation;
and the number of embedding changes that is a quantity
closely related to the length of the embedded data.
Given two embedding schemes that share the first three
attributes, the scheme that introduces fewer embedding
changes will be less detectable. The rest of this section
reviews some of the existing proposed steganography
methods.
DCT domain embedding techniques are very
popular due to the fact that JPEG which is a DCT-
based image format is widely used in the public domain
in addition to being the most common output format of
digital cameras. Some of steganographic embedding in
DCT domain are Outguess [3],
F5 [4], model-based
[5], perturbed quantization (PQ) [6], and Matrix
embedding [7].
A cover selection method [9,10], like an image
retrieval method, retrieves images based on their fitness
to carry a given secret image. Cover selection problem
was studied in [10] by investigating three scenarios in
which the embedder has either no knowledge, partial
knowledge, or full knowledge of the steganalysis
technique. The main idea in [9] is based on dividing the
secret image into blocks of size 4×4 where for each
secret block, the most similar block in the host image is
IEEE 8th International Conference on Computer and Information Technology Workshops
978-0-7695-3242-4/08 $25.00 © 2008 IEEE
DOI 10.1109/CIT.2008.Workshops.34
379

found and the secret block is placed there. The host
image is found from an image database, in such a way
that it has the most number of similar blocks to those of
given secret image. To find the similarity between
blocks, they used texture analysis measures based on
Gabor filter. Then, the location addresses of the blocks
in host image which are replaced by blocks of secret
image are saved. Then, this data is converted to a bit
string and coded by Hamming code. This bit string is
embedded in determined DCT coefficients of the
modified host image and the blocks for embedding are
selected using a key which is the seed of a random
sequence generator. One of the advantages of this
method is increasing the embedding capacity by hiding
only blocks indices in DCT coefficients of host image.
One of the difficulties in texture based similarity
measure presented in [9] is that, they compare only the
content of two 4×4 blocks without considering the
effect of pixels in close neighborhood to the blocks.
Therefore, by replacing similar blocks with each other,
virtual edges may appear in borders and corners of a
replaced block.
In this paper to remove the blocking effect and
hence, improve the quality of stego images, we used
the neighborhood information beside block texture
information. Each block in the secret image is taken as
a texture pattern for which the most similar block is
found among the blocks of the host image. The
embedding procedure is carried on by replacing these
small blocks of the secret image with blocks in host
image in such a way that least distortion is imposed.
We have used block texture combined with
neighborhood information to measure the similarity of
blocks. Our experimental results showed a high level of
capacity and minimum distortion on images. In this
way, we present a method for image hiding that uses
the concept of block similarity between host and secret
images. We used texture information to obtain a mean
for summarizing the information of each image. Also
block neighborhood information is used to avoid
generating virtual edges and furthermore, K-means
algorithm is applied to improve the speed of finding the
best host image. The stego and restored secret images
have high quality and we verified that, using recent
powerful blind steganalyzers, one can not discriminate
between clean and stego images reliably.
In section 2, we introduce the proposed
steganographic method for gray level images.
Performance of the proposed technique is analyzed in
Section 3 and finally, section 4 outlines some
concluding remarks.
2. Proposed steganography method
Given a secret image, the proposed steganography
algorithm finds the best candidate host image from an
image database. Selecting the best cover is based on
the similarity of secret and cover image blocks. Like
the method which was proposed in [9], the main idea is
to find texture pattern blocks of the secret image, in the
host image, and save their addresses in host image.
Some statistical features of a block and its neighbor-
hood information are used to measure similarity of
image blocks. The structure for the proposed
steganography algorithm is shown in Figure 1.
2.1. Feature Extraction
Image properties such as image texture could be
used to categorize the images. The crude measures of
image texture would be the mean, variance, and
skewness of image blocks which are simple and can
efficiently be computed. In order to compute the
similarity between secret and cover images blocks, all
the images are divided into blocks of size 4×4. For
each block, the statistical values such as mean,
variance, skewness of 2×2 sub-blocks and, the
neighborhood information, as shown in Figure 2, are
calculated.
Four 2×2 sub-blocks are considered in up-right, up-
left, down-right and, down-left corners in a 4×4 block.
Figure 1. Selecting the best host image to cover a given secret image.
Feature
Extraction
Cover Image Database
Feature
Vectors
Cluster
Prototypes
4×4
Blocking
Feature
Extraction
Clustering
Secret
Image
4×4
Blocking
Selecting Best Host Image
Image
Database
Selected
Host Image
380

The average intensity value of four pixels that are
adjacent to each side of a 4×4 block, are computed and
considered as neighborhood information of that block
side. In this way, a 16 dimensional feature vector is
obtained (12 statistical values and 4 neighborhood
mean values). By searching a host image from image
database, the image which provides the best similarity
to the secret image will be selected.
2.2. Finding the best host image
Due to the existence of large number of feature vectors
for each image in the database an efficient method must
be used for indexing and searching the blocks.
Therefore, K-means is applied to each image of the
database to cluster feature vectors. For selecting a
block of host image, similar to a secret image block,
the feature vector extracted from the secret image is
compared to the cluster indices. The most similar block
is found from the cluster whose cluster prototype is the
nearest to the selected secret block feature vector.
For each block in secret image, its feature vector is
calculated and compared to cluster prototypes in
database. Each cluster prototype indicates the index of
a group of similar blocks. Having determined the most
similar cluster prototype, the most similar block is
searched in that cluster and is selected as the best
choice in this group. Euclidean distance is then applied
to measure the closeness of image blocks in database to
the blocks of secret image. This procedure is carried on
for all blocks of secret image and finally, the image
with the most number of similar blocks in database is
chosen to be the host image.
2.3. Encode and decode a secret image
After host image selection, we used the same approach
as in [9] to embed the secret image to the selected host
image that is as follows: Each block of secret image is
replaced with the most similar block to it in host image.
The positions of secret image blocks in host image are
saved. In next stage, the sequence of mentioned block
positions is changed to a bit string. Then a seed for
generating a random sequence of location addresses
(key) is considered.
Figure 2. Regions for computing features.
Figure 3. Middle DCT coefficients of an 8×8 block.
The position address bit string will be hided in middle
DCT coefficients of host image.
The way of doing steganography in the DCT
domain is to modulate the relative size of DCT
coefficients. The algorithm is described in [11] as
splitting the image into 8×8 blocks and calculating the
DCT of each block. Then two middle-frequency
coefficients are chosen and agreed upon by both send
and receive parties. A block encodes a 1 if
DCT(a,b) >
DCT(c,d)
and 0 otherwise.
In the encoding step, the coefficients are swapped if
their relative size does not match with the bit to be
encoded. Determined DCT coefficients are shown in
Figure 3.
The sender sends the modified stego image SI to the
recipient. Decoding is straightforward because the
recipient first forms the random sequence by using the
same key K and then, retrieves the embedded bit string
from the DCT coefficients of the blocks. The extracted
bit string is simply Hamming decoded and then indices
of the secret image blocks are extracted. Through the
knowledge of block indices, the secret image can be
reconstructed.
2.4. Steganalysis
Each steganalyzer is composed of feature extraction
and feature classification components. In this context,
two techniques which are studied in this work, take
distinct approaches in obtaining distinguishing statistics
from images. One of these techniques is wavelet-based
steganalysis (WBS) proposed by [12,13]. In feature
extraction part of this method, statistics such as mean,
variance, skewness, and kurtosis are calculated from
each wavelet decomposition subband. Additionally, the
same statistics are calculated for the error obtained
from a linear predictor of coefficient magnitudes of
each subband. Feature-based steganalysis (FBS) [17]
obtains a set of distinguishing features from the DCT
and spatial domains. It is shown in [16] that FBS
technique outperforms other techniques such as WBS
and binary similarity measures.
381

3. Experimental results
To examine the proposed method, different
experiments were done. Peak signal-to-noise ratio
(PSNR) given by the following formula is employed
for evaluation purposes:
MSE
L
PSNR
2
10
log10=
(1)
∑
=
−=
N
i
ii
xx
N
MSE
1
)(
1
(2)
L is the maximum gray level value and
i
x is the real
value of the pixel with
i
x being the pixel value after
modification and finally, N is the total number of
pixels.
3.1. Image database
We obtained a list of 330 JPEG images from some
typical images and some random ones from
Washington University image database [16]. The image
database contained cover images of size 256×256
where the secret images are of size 64×64. All images
were converted to grayscale and saved with quality
factor of 75. Each secret image was used to obtain a
stego dataset using the proposed technique.
3.2. Average capacity and Time complexity
evaluation
Fridrich in [17] showed that the average
steganographic capacity of grayscale JPEG images with
quality factor 70 is approximately 0.05 bits per non-
zero DCT coefficient. In the proposed method, each
64×64 size secret image has 256 blocks of size 4×4.
For embedding this secret image in a host image, we
should find 256 corresponding similar blocks. If each
host image size is 256×256, it has 4096 blocks of size
4×4. Therefore, for addressing 4096 blocks, we need
12 bit addresses. After applying Hamming code
algorithm, for each 4 bits, 7 bits, and totally, for
addressing each block 21 bits and for the whole 256
blocks, 5376 bits are needed. In this way, a 64×64
secret image is embedded in a 256×256 host image and
the embedding rate is 0.06 per bit. Experimental results
are carried out on a 2046 MB PIV processor using
MATLAB 7.1 and image processing toolbox 5.0.2. It
should be noted that MATLAB codes are usually 9 or
10 times slower than their C/C++ equivalents [18].
Table 1 shows the results of the time evaluation of the
proposed method.
Table 1. Average execution time.
Stage
Selecting Best
Host Image
Decoding
Time
Execution Time
4 Minutes 1 Second
3.3. Performance of selecting a host image
In this evaluation, we considered capacity,
perceived quality, and undelectability that are
presented in Figure 4. The figure shows the
performance of the proposed method on some
randomly selected secret images. Figure 4(b) is the best
candidate for the given secret image. Figure 4(c) and
4(d) show the host image in which the secret image is
hidden and the restored secret image, respectively.
From Figure 4(c), we can see that the quality of the
stego image is high, and unintended observers will not
be aware of the existence of a hidden image in it.
Indeed, it is impossible to distinguish between Figure
4(b) and (c) or between Figure 4(a) and (d) using naked
eye. This indicates that the value and normal usage of
the secret image are preserved. The average of PSNR
of the stego images is more than 39 dB which shows
that the cover images have excellent imperceptibility
after the secret image is embedded in them.
Table 2 shows the effect of using only block texture
information as suggested by [9] and using
neighborhood information as suggested by our method.
The results show that using neighborhood information
increases the quality of stego and restored secret image.
3.4. Steganalysis
In this section, we evaluate the security of the proposed
algorithm using two blind steganalyzers that use
features constructed in wavelet domain (WBS), and
features calculated in the DCT domain (FBS). In WBS,
a Fisher Linear Discriminator (FLD) and in FBS a
nonlinear Support Vector Machine (SVM) is trained to
discriminate clean and stego images. One hundred
images from database were chosen randomly for
testing, while the remaining images were used for
training. This partitioning was repeated a total of ten
times, with different random subsets used for training
and testing each time. For each of the ten partitions, the
SVM/FLD was trained with the statistics from the
training image subset. Finally, the trained classifier was
tested against the previously unseen images. The
average of detection accuracy is shown in Table 3.
As can be seen, the proposed method with payload
of approximately 0.067 bits per cover image bit can not
be reliably detected by any of the two steganalyzers.
382

Table 2. Comparison of stego and restored secret image PSNR with and without
considering neighborhood information.
Considering
neighborhood information
Average
Stego Image PSNR
Average Restored Secret
Image PSNR
No 36 34.5
Yes 39.5 35
Table 3. Blind steganalyzers detection accuracy.
Average Secret
Image Size
Average Cover
Image Size
Steganalyzer
False
Positives
True
Positives
Detection
Accuracy
WBS(FLD) 37.11% 53.77% 58.33%
4.3 KB 63.5 KB
FBS (Non-linear SVM) 44.6% 37.2% 46.3%
Secret Image(a)
Selected cover image
(b)
Stego Image(c)
Extracted secret
Image (d)
Size(KBytes)=4.25 Size(KBytes)=64.6 PSNR=39.50 PSNR=35.51
Size(KBytes)=4.26 Size(KBytes)=64.8 PSNR= 41.31 PSNR= 37.23
Size(KBytes)=4.25 Size(KBytes)= 64.6 PSNR= 39.62 PSNR= 36.36
Size(KBytes)=4.14 Size(KBytes)= 64.8 PSNR= 39.51 PSNR= 34.14
Figure 4. (a) Secret image. (b) Selected cover image. (c) PSNR Between the stego and cover
image. (d) PSNR between the original and extracted secret image (Units are in dB).
383

4. Conclusion
In this paper, we proposed a data hiding scheme that is
imperceptible while a big secret image is concealed in
a cover image. The main idea is based on dividing the
secret image into blocks and considering these blocks
as units for embedding. Then, using the similarity
measure, provided by feature vector, the most similar
block in the host image is found and the entire secret
block is replaced there.
Considering the neighborhood information of
blocks, prevents the method presented in [9] to
outbreak the virtual edges in sides and corners of
replaced blocks. The main achievements of the
proposed steganography method are: (i) reduction of
the host image distortion, and (ii) increased security. In
addition our method benefits from the advantages of
increasing the embedding capacity by hiding only
blocks indices (location addresses) in DCT coefficients
of host image as suggested in [9].
Our approach aims to reduce the risk of detection,
while keeping a high embedding capacity. This gain is
more important for long messages than for shorter ones
because longer messages are, in general, easier to
detect. We showed that cover selection method
provides good embedding efficiency and its relative
embedding capacity densely covers the range of large
payloads, making it suitable for practical applications.
The experimental results show that applying the
steganalysis methods on stego images can not reliably
detect stego and clean images.
5. References
[1] N.
Johnson, S. Jajodia, “Steganalysis of images created
using current steganography software”, in Proc. Int.
Workshop on Info. Hiding, Germany, 1998, pp. 273-289.
[2] G. Greg, “Steganalysis Gets Past the Hype”, IEEE
Distributed Sys. Online, April, 2005, vol. 6, no. 4, pp. 1-5.
[3] N. Provos, “Defending against statistical steganalysis”, in
Proc. 10th USENIX Security Symp., 2001.
[4] A. Westfeld, “F5-a steganographic algorithm: high
capacity despite better steganalysis”, in Proc. 4th Int.
Workshop on Info. Hiding, 2001.
[5] P. Sallee, “Model-based steganography”, in Proc. Int.
Workshop on Digital Watermarking, Seoul, Korea, 2003.
[6] J. Fridrich, M. Goljan, D. Soukal, “Perturbed
quantization steganography with wet paper codes”, in Proc.
ACM Multimedia Workshop, Germany, 2004.
[7] J. Fridrich, D. Soukal, “Matrix Embedding for Large
Payloads”, IEEE Trans. on Info. Forensics and Security, vol.
1, No. 3, 2006.
[8] P. Su, C. Kuo, “Steganography in JPEG 2000
compressed images”, IEEE Trans. Consum. Electron. 2003,
Vol. 49 Issue 54, 824–832.
[9] Z. Kermani, M. Jamzad, “A Robust Steganography
Algorithm Based On Texture Similarity Using Gabor Filter”,
IEEE Int. Symp. on signal processing and Info. Technology,
2005, pp. 578-582.
[10] M. Kharrazi, H. Sencar, N. Memon, “Cover Selection
for Steganograpic Embedding”, in Proc. ICIP, 2006, pp.117-
121.
[11] S. Katzenbeisser, F. Petitcolas, “Information Hiding
techniques for steganography and digital watermarking”,
2000, ISBN 1-58053-035-4.
[12] S. Lyu, H. Farid, “Detecting hidden messages using
higher-order statistics and support vector machines”, in Proc.
5th Int. Workshop on Info. Hiding, 2002.
[13] S. Lyu, H. Farid, “Steganalysis using color wavelet
statistics and one-class support vector machines”, in Proc.
SPIE 5306, 2004, pp. 35–45.
[14] J. Fridrich, “Feature-based steganalysis for jpeg images
and its implications for future design of steganographic
schemes”, in Proc. 6th Info. Hiding Workshop, Toronto,
2004.
[15] M. Kharrazi, H. Sencar, N. Memon, “Performance study
of common image steganography and steganalysis
techniques”, Journal of Electronic Imaging 15(4), 041104,
2006.
[16]http://www.cs.washington.edu/research/imagedatabase/g
roundtruth/
[17] J. Fridrich, T. Pevny, J. Kodovsky, “Statistically
Undetectable JPEG Steganography: Dead Ends, Challenges,
and Opportunities
”, MM&Sec’07, Dallas, USA.ACM, 2007.
[18] K. Cho, J. Jang, K. Hong, “Adaptive skin-color filter”,
Pattern Recognition, 2001, Vol.34, Issue 5, pp. 1067-1073.
384