2010 International Conference on Multimedia Information Networking and Security

Image Steganogaraphy based on Complex Double Dual Tree Wavelet Transform

Sushil Kumar
Department of Mathematics,
Rajdhani college,University of Delhi,
New Delhi,India
e-mail: [email protected]

S. K. Muttoo
Department of Computer Science,
University of Delhi,
Delhi, India
e-mail: [email protected]
Abstract— This paper investigate the applications of Dual Tree steganography such as DCT. DCT remained famous domain
Complex Wavelet Transforms to steganography and compares of embedding as it was the major compression technique in
it with other transform domains such as DWT, Wavelet-like earlier JPEG compression technique. Wavelet transform
transform domain, viz, Slantlet domain and Double Density replaced the DCT in the JPEG 2000 and also found to be
Dual Tree DWT domain. It is shown that Dual Tree Complex better option than DCT in terms of embedding capacity and
Wavelet Transform domain is better option than others as it robustness. However, DWT lacks directional selectivity for
offers better imperceptibily and high embedding capacity. It is diagonal features and shift invariance. The dual tree complex
also a well known fact that Complex Wavelet Transform not wavelet transform (DT-CWT) has a modest amount of
only provides shift invariance but also offers better directional redundancy, but it provides shift invariance and good
selectivity over the discrete wavelet transform. We present a directional selectivity. Kingsbury [6] introduced Dual-Tree
steganographic algorithm in four transform domains as Complex Wavelet Transform (DTCWT) and Selesnick [13-
mentioned above. The original string message is first 14] provided Double-Density DWT (DD DWT) and Dual-
converted into binary cell array with the help of self- Tree DWT (DT DWT) for complex analysis. They have
synchronizing variable length codes, viz., T-codes before the shown that DT-CWT has a modest amount of redundancy, but
embedding. We then embed the secret message in the it provides shift invariance and good directional selectivity.
insensitive middle and high sub-bands obtained from the cover The 2-D DT-CWT is based on two distinct scaling functions
image after applying these transforms on it. To see the visual and two distinct wavelets, whereas Double-Density DWT is
effect of data embedding in the Complex Wavelet Transform based on a single scaling function (low pass) and two distinct
domain, we use the most convenient and simple modified LSB wavelets (high pass). On the application of 2-D DD DWT , the
approach. The metric used for evaluating the visual quality of cover image is decomposed into nine sub-bands labeled LL,
stego-image is PSNR. The experimental results show that the LH1, LH2, H1L, H1H1, H1H2, H2L, H2H1, H2H2, respectively.
dual tree complex wavelet transform is better option than the The Double Density Dual-Tree DWT (DD DT DWT)
Wavelet, Wavelet-like transform, i.e., Slantlet transform and possesses simultaneously the properties of DT DWT and DD
Double Density Dual Tree DWT for data hiding in terms of DWT. The structure of DD DT DWT consists of two
visual quality and embedding capacity, though poor to oversampled iterated filter-bank that operate in parallel. Each
external attacks such as Gaussian. row of the cover image is first subjected to 1-D DD DT DWT
Keywords- DWT, SLT, DD DT-CWT, PSNR, BPP, JPEG 2000 decomposition in which one is real decomposition that uses
the real component of DD DT DWT and the other is an
I. INTRODUCTION imaginary decomposition. After the 1-level decomposition we
obtain four times wavelets than the ordinary 2-D DD DWT
The discrete wavelet transform (DWT) have been used as a decomposition.
popular tool to embed the message in the frequency In this paper we propose four high capacity data hiding
coefficients of high sub-bands obtained on applying it to the techniques based on the transforms, viz., Wavelet (wavcdf97),
original cover image. There are applications that demand high Slantlet, Double Density Dual Tree DWT and Double Density
capacity, e.g., in the advance research on network Dual Tree Complex Wavelet, respectively. We embed data
transmission security and lossless embedding such as in into insensitive Wavelet/Slantlet/DT-CWT sub-bands using
military, legal and medical imaging domains. The integer the modified LSB method. The embedding capacity can reach
wavelet transform provides better embedding capacity and to ¾ of the cover image (grayscale) in Wavelet and Slantlet
better imperceptibility than the earlier used transforms in domains and 6/7 and more of the cover image (grayscale) in

978-0-7695-4258-4/10 $26.00 © 2010 IEEE 93

DOI 10.1109/MINES.2010.28
DD DT_CWT domain. The proposed algorithms are III. PROPOSED ALGORITHM
implemented in Matlab 7.0 using 256 x 256 size images. We propose DD DT DWT/DD DT-CWT based embedding
Experimental results shows that the DD DT-CWT method has algorithm in which the secret message is embedded in the
not only higher capacity but also better visual quality than insensitive frequency coefficients using modified LSB
other three methods, viz., Wavelet (wavcdf97), Wavelet-like approach. LSB steganography is the simplest steganographic
transform (Slantlet) and DD DT DWT. These high capacity technique used in popular steganographic tools such as S-
steganogrphic techniques can be applied to e-government, e- Tools, Steganos and StegoDos, where embedding is done in
business, e-law enforcement, military system and e-medical the spatial domain [5]. The sequential LSB has a serious
system. security problem whereas random or modified LSB in which
The rest of the paper is organized as follows. The Wavelet, secret message can be randomly scattered in stego-images
Wavelet-like transform and DD DT-CWT are introduced in provides an improvement over the steganographic security.
Section II. The proposed algorithm is summarized in Section The basic idea of LSB embedding is to embed the message bit
III, while the experimental results are presented in Section IV. at the rightmost bits of pixel value so that the embedding
Conclusion and future scope are presented in Section V. method does not affect the original pixel value greatly. The
formula for the embedding is as follows: x’ = x - x mod
II. EMBEDDING DOMAIN 2k + b , where k is the number of LSBs to be substituted. The
extraction of message from the high frequency coefficients is
In designing steganographic algorithm, special consideration given as: b = x mod 2k
is given to the domain in which the embedding will take place.
No doubt , frequency domain have shown to be better option The proposed algorithm proceeds as follows:
than the spatial domain for steganographic algorithms used so
far, for high capacity, better visual quality and robustness.
Some transforms domains used include Discrete Fourier Step 1. Obtain the secret data by applying best T-codes to the
given input text/message. An encoded key is generated.
Transform ( DFT), Discrete Cosine Transform (DCT),
Step 2. Decompose the cover image low and sub-bands ( viz.,
Discrete (Haar/cdf(2,2)) Wavelet Transform (DWT), LL, HL, LH and HH) by applying 2-D
Directional transforms such Curvelets, Contourlets etc and DWT/SLT/DD DT DWT/DT-CWT.
Wavelet-like Transform such as Slantlet Transform (SLT). Step 3. Approximate/Normalize the frequency coefficients to
Panda et al. [11] have shown that SLT possesses better energy integers using a threshold value, usually 0.8.
compaction properties than DWT. Sushil Kumar and Step 4. Embed secret data in the middle and high frequency
bands (LH, HL and HH) using the modified LSB method.
S.K.Muttoo [9, 16] have shown that SLT further provides a
much better approximation of the Human Visual System Step 5. Obtain the stego-image by taking the inverse
(HVS) than the DWT. However, DWT and SLT have their transform to the modified image.
inability to differentiate between opposing diagonal features,
i.e., have poor directional selectivity and lack of shift
invariance. Selesnick [15] has shown that Complex Wavelet
Transform (CWT) has the advantage of excellent shift The extraction method is the reverse method of embedding
invariance at the cost of 2:1 redundancy for 2-D signals. The algorithm. It consists of the following steps:
DT DWT not only overcomes the problem of poor directional
selectivity in DWT, but can also discriminate between
opposing diagonals with six different sub-bands oriented at Input: stego-image, encoded-key, stego-key
Output: original message
150, 750, 450, -150, -750, and – 450.
Step 1. Decompose the stego-image into low and high sub-
The major requirement of any steganographic algorithm is bands using the transforms Wavelet/Slantlet/DD DT DWT/DT
imperceptibility. So it is necessary to limit the distortion CWT.
applied to each individual coefficient of the transform Step 2. Extract the secret message using the stego-key used in
decomposition. One of the solutions may be to find the embedding technique
encoding scheme that minimizes the bit conversion while Step 3. Obtain the original message by decoding the secret
embedding the secret message. Other solution may lie in using message using T-decoding algorithm and encoding key.
visual tests such as Just Noticeable Distortion (JND) profile.
In the proposed algorithm we have used Self-synchronizing
variable length codes, T-codes which have shown to be a
better encoding scheme than a popularly known Huffman’s IV. EXPERIMENTAL RESULTS
coding [18,20].

94
The proposed algorithm is implemented in four domains, viz.,
Wavelet domain, Slantlet domain, DD DT DWT and DT- 70
CWT domain. We have used 256 x 256 size images and
results are obtained by running the algorithm in Matlab 7.0
software. The summary of the results obtained for some of the
images is given in table 1. Figure 1 shows the comparison 65
between four transforms in terms of PSNR values vs BPP (bits
per pixel) rate. The results obtained show that DT-CWT
outperforms the other transforms in terms of visual quality and
60
embedding capacity.

ps nr
TABLE 1. COMPARISON OF DWT, SLT, DD DT DWTAND DT-CWT
METHODS(MESSAGE LENGTH=6000) 55

Image Wavelet 2-D 2-D DD 2-D DT

(wavcdf97) Slantlet DT-DWT CWT
PSNR PSNR PSNR PSNR 50
(db) (db) (db) (db)
I1 44.350043 53.601686 60.445225 63.436141
I2 44.134947 53.553464 60.493630 63.094094 45
I3 44.329890 53.524250 60.962041 63.672915 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
BPP
I4 44.489671 53.570887 60.587625 63.406725
I5 46.074744 53.603920 60.089842 65.069457 Figure 1: Comparison of Wavelet (blue/bottom line), Slantlet (
I6 46.528802 53.293788 61.391552 63.898702 green/second line form bottom), Double Density Dual Tree Discrete
I7 46.726047 53.629027 61.219253 63.934244 Wavelet transform (red /third line from bottom), and Double Density
(I1. Peppers.png, I2. Tooth1.png, I3. Twins.small.jpg, I4. Dual Tree Complex Wavelet transform ( Top line) used in embedding
Cameraman.tif, I5. Zoneplate.png, I6. Arts.png, I7. Lena.png) algorithm based on LSB method in terms of psnr vs bpp with image
=’lena.jpg’

The table 2 shows that the LSB method is nonrobust to attacks

such as Gaussian etc. We would like to say here that the three V. CONCLUSION AND FUTURE SCOPE
main requirements of any steganographic algorithm are
Imperceptibilty, Capacity and undetectabilty (i.e.security from the A steganographic algorithm based on three different
third party). Robustness, however, can be limited to compression transforms , DWT, SLT and DT-CWT using the
and channel noise only. modified LSB technique is presented and compared in
terms of visual quality and embedding capacity. The
Table 2: Gaussian effect (= 0.0001) to added to stego-images
modified LSB method can be implemented conveniently
and for high capacity data can be embedded using two or
more LSBs bits per pixel. The disadvantage of the LSB
method is that it is non-robust. Cleik et al. [21] have
proposed Generalized-LSB data embedding that offers
finer grain scalability along the capacity distortion curve.
The purpose of this paper was to investigate Complex
double dual-tree wavelet as an application to
steganography and compare it with already known
transform domains such as Wavelets and Wavelet-like
transforms. The authors have also applied other
techniques such as Thresholding technique [16] and
Wavelet-fusion technique [19] in place of LSB and the
results obtained thereof also proves our claim that DT-
CWT domain is superior than DWT or SLT domain in
view of visual quality of stego-image and embedding
capacity. Thompson et al. [18] have shown that DT-CWT
domain is good robustness against several attacks such as

95
 Wiener filtering, Median filtering, Mean filtering, JPEG [13] Selesnick I. W., “The double density DWT”, in: Wavelets in
compression, and AWGN attacks for image watermarking Signal and Image Analysis: From theory to Practics, A. Petrosian
based on Spread Spectrum and Quantization Index and F. G. Meyer, (Eds), Norwell, MA: Kluwer, 2001.
Modulation. We are working on finding a secure high
capacity steganographic algorithm in the DT-CWT domain [14] Selesnick I. W.,“The Double-density dual-tree DWT”, IEEE
that is robust to common attacks such as low-pass filtering, Trans. On Signal Processing, 52(5) (2004), 1304-1314.
JPEG compression and Gaussian noise.
[15] Selesnick I. W., R. Baraniuk, and N.G. Kingsbury, “The
dual-tree complex wavelet transform: A coherent framework for
multiscale signal and image processing”, IEEE Signal Proc.
Magazine, (2005) 123-151.
REFERENCES
[1] Alpert B., Coifman G.R., and Rokhlin V., “Wavelet-likebases [16] Sushil Kumar and S.K. Muttoo, “Distortionless data hiding
for the fast solution of second kind integral equations”,SIAM using Slantlet transform”, MINES 2009, Wuhan, China
J.Sci. Comput., 14: 159-184.http:///dx.doi.org/10.1137.0914010
[17] Thompson A. I. et al., “Watermarking for Multimedia
[2] Celik Mehmut Utku et al., “Lossless Generalized-LSBData Security using Complex Wavelets”, Communicated for
Embedding”, IEEE Transactions on image processing,vol. 14, publication.
No. 2, Feb. 2005.
[18] Titchener, M.R., “Generalised T-codes: extended
[3] Do M.N. and Vetterli M., “The Contourlet transform: An construction algorithm for self- synchronization codes”, IEEE
efficient directional multiresolution image representation”, IEEE Proc. Commun., Vol. 143, No.3, pp. 122-128, 1996
Trans. On Image Processing, 14(12), 2091-2106.
[19] Tolba, M.F.; Ghonemy, M.A.-S.; Taha, I.A.-H.; Khalifa,
[4] Joachim J. Eggers, R. Bauml and Bernd Girod, “A A.S., “High Capacity Image Steganography using Wavelet-Based
Communication approach to Image Steganography”, Proceedings Fusion” , Computers and Communications, 2004, Proceedings.
of SPIE., Security and Watermarking of Multimedia Contents IV, ISCC2004. Ninth International Symposium on Volume 1, Issue,
Vol. 4675,San Jose, CA, Jan., 2002. 28 June-1 July 2004, Vol.1, 430-435

[5] Johnson, N.F. and S. Jajodia, “Exploring steganography: [20] Ulrich G., “Robust Source Coding with Generalised T-
Seeing the unseen”, IEEE Computer, 31: 26-34, 1998 codes”, a thesis submitted in the University of Auckland, 1998.

[6] Kingsbury N.G.,“Image processing with complex wavelets”, [21] Zhang Tao and Xijian Ping, “A new approach to reliable
Philos. Trans. R. Soc. London A, Math. Phy. Sci., 357(1760) detection of LSB steganography in natural images”, Signal
(1999) 2543-2560 Processing, 83 (2003), 2085-2093,
www.ComputerScienceWeb.com
[7] Maitra M. and Chatterjee A., “A Slantlet transform based
intelligent system for magnetic resonance brain image
classification”, Biomedical Siganl Processing and Conrol, 1, pp.
299-306

[8] Muttoo S.K. and Sushil Kumar,“Data Hiding in JPEG

images”, BVICAM’s IJIT, Vol. 1, No.1, Jan.-July, 2009.

[9] Muttoo S.K. and Sushil Kumar, “Secure image Steganography

based on Slantlet transform”, ICM2CS-09, JNU, New Delhi,
India, 2009

[10] Navas K.A., Mathew N.M.,and Sasikumar M. “Contourlet

based data hiding in medical images”, CSI, 2003.

[11] Panda G. and Meher S.K., “An efficient approach to signal

compression using slantlet transform”, IETE Journal of Research,
Vol. 46, No. 5, September, pp. 299-307, 2000

[12] Selesnick I. W., “The Slantlet Transform”, IEEE

transactions on signal processing,Vol. 47, No. 5, May,1998, pp.
1304-1312.

96