In this paper, we propose a discrete wavelet transform. (DWT) based watermarking method, which can be conveniently integrated in up-coming JPEG-2000 ...
A SECURE WATERMARKING FOR JPEG-2000 Yong-Seok Seo, * Min-Su Kim, *Ha-Joong Park, *Ho-Youl Jung, *Hyun-Yeol Chung, **Young Huh, **Jae-Duck Lee VR Center, ETRI, Yusong P.O. Box 106, Taejon, 305-600, KOREA *School of EECS, Yeungnam Univ., Kyungsan, 712-749, KOREA **KERI, P. O. BOX 20, Changwon, 641-120, KOREA ABSTRACT In this paper, we propose a discrete wavelet transform (DWT) based watermarking method, which can be conveniently integrated in up-coming JPEG-2000 baseline. Conventional DWT based watermarking techniques insert a watermark into the coefficients after the transform completed, while the proposed method inserts a watermark into the coefficients obtained from ongoing process of lifting for DWT. The proposed method allows us to determine selectively frequency characteristics of the coefficients where watermark is embedded, so that the watermark cannot be easily removed or altered even when filter-banks for DWT was known. Through the simulations, we show that the proposed method is more secure and more robust against various attacks than the conventional DWT based watermarking.
1. INTRODUCTION Recently, many research results have claimed the robustness of the frequency domain based watermarking compared to the spatial domain based one [1]. In addition, the frequency domain methods have some advantages because most of the signal processing operations can be well characterized in the frequency domain, and many good perceptual models are developed in the frequency domain [1-5]. In particular, discrete cosine transform (DCT) based watermarking [1,2] have been widely used, since it has been well matched the current image/video compression standards such as JPEG, MPEG 1-2, etc. Considering new image compression standards such as JPEG-2000 [7,8], however, DWT will be very attractive transform for watermarking. By embedding a watermark in the same domain as the compression scheme used to process the image, we can anticipate lossy compression because we are able to anticipate which the transformed coefficients will discard by the compression scheme [9]. This makes the watermark to be robust against lossy
compression. In addition, it is possible to integrate the watermarking process into JPEG-2000. A unified algorithm will be useful for many applications that require the compression and the watermarking simultaneously, such as digital still camera, web based monitoring camera. Conventional DWT based watermarking techniques which insert a watermark into transformed coefficients after completed the transform are not sufficient to integrate straightforwardly into JPEG-2000, because the inserted watermark can be easily altered by using the revealed filter-banks for compression (the same filter-banks is used for watermarking). In this paper, we present a new watermarking method, which inserts a watermark into the coefficients obtained from ongoing process of lifting for DWT. This scheme allows convenient integration of watermarking process in JPEG-2000 and is more secure and robust than conventional DWT based watermarking methods [2-5]. Section 2 offers a brief overview of the JPEG-2000 baseline, mainly lifting for DWT. The lifting algorithm will be basically applied to develop the proposed watermarking in section 3. We also show that the method can be easily integrated with slight computational complexity in the JPEG-2000 baseline. In Section 4, we present our experimental results and discussions. Finally, Section 5 contains the conclusions. 2. LIFTING BASED DWT IN JPEG-2000 JPEG-2000 baseline employs both reversible integer-tointeger and non-reversible real-to-real filter-banks for DWT [7,8]. Lifting will be applied to DWT for its low computational complexity and low memory requirements. Here, we focus on the lifting based Daubechies 9-7 taps filter-banks [6], as it is default filter-banks for lossy JPEG2000 baseline and will be employed to develop the proposed watermarking method. Fig. 1 illustrates the lifting for the 9-7 taps filter-banks, which consists of four lifting steps and one scaling step. This procedure can be written as (1), each of which corresponds to respective columns of Fig. 1. For a given input sequence xi, yi (for i =
x0 x1
x0 x1
b0
c0
d0
a1
b1
c1
d1
a2
x2
b2
a3
x3 α
x4
a0
α
x5
a4 a5
β β
x7 x8 x9 [step 1]
b4 b5
d2
c3
b3
a6
x6
c2
γ γ
b6
d3 δ
c4
δ
c5
d4 d5
a7
b7
c7
d7
a8
b8
c8
d8
a9
b9
c9
d9
[step 2]
[step 3]
[step 4]
x2
y1
x3
y2
x4
y3
x5
y4
x6
1/ K
y5
K
d6
c6
y0
y6
y8
x10
[step 5]
= x + α ⋅ ( x i −1 + x i +1 ), for i = 2n + 1 ai i xi , otherwise =
(1-1)
= a + β ⋅ (a i −1 + a i +1 ), for i = 2n bi i ai, otherwise =
(1-2)
= b + γ ⋅ (b i −1 + b i +1 ), for i = 2n + 1 ci i bi , otherwise =
(1-3)
= c + δ ⋅ (c i −1 + c i +1 ), for i = 2n di i ci , otherwise =
(1-4)
= yi =
(1-5)
K ⋅ di , 1/ K ⋅ di ,
for i = 2n + 1 otherwise
x8 x9
Fig. 1. Lifting for 9-7 taps forward DWT.
Where n∈Z and parameters, α, β, γ, δ, Κ, are given in [7,8]. Its inverse DWT is performed by the reverse procedure of (1). 3. LIFTING BASED WATERMARKING
Once, we could consider a method of inserting a watermark into the coefficients obtained at any step of the lifting in Fig. 1. However, this method does not ensure transparency of watermark, because all intermediate coefficients, ai, bi, ci, possess low-frequency components. For the transparency, coefficients should be high-pass filtered version of an input signal [1].
α
x7
y7
y9
α
[step 1]
Ω0
a0
b0
a1
b1
a2
b2
a3
b3
a4 a5
β β
b4 b5
Ω1 Ω2 Ω3
ω γ−ω
a6
b6 γ−ω
a7
b7 ω
a8
b8
a9
b9
a10
b10
[step 2]
watermark insertion position
2n + 1) and yi (for i = 2n) obtained in the last step represent high-pass and low-pass filtered coefficients, respectively.
Ω4
y0
c1
d1
y1
c2
d2
c3
d3
c5
δ δ
d4 d5
Ω6 ω −ω c6
d6
c7
d7
c8
d8
c9
d9
c10
d10
Ω7
Ω8 Ω9 Ω10 [step 3]
d0
c4
ω −ω
Ω5
c0
[step 4]
[step 5]
y2 y3 1/ K
y4
K
y5 y6 y7 y8 y9 y10
[step 6]
Fig. 2. A modified lifting for 9-7 taps forward DWT. Here, watermark is inserted in the coefficients, Ωi.
Here, we propose a new watermarking which inserts a watermark into the coefficients obtained from ongoing process of lifting, maintaining the transparency. Moreover, the proposed method could alter frequency characteristics of the coefficients where watermark is inserted. Fig. 2 shows the main idea, which consists of five lifting and one scaling steps. Watermark is embedded into the coefficients, Ωi (for i = 2n + 1), obtained in [step 3]. [step 3] and [step 4] can be written as
= b i + ( γ − ω) ⋅ (b i −1 + b i +1 ) Ωi + ω ⋅ (b i−3 + b i+3 ) , for i = 2n + 1 (2-1) = b i , otherwise = Ω i + ω ⋅ (Ω i −1 + Ω i +1 ) ci − ω ⋅ (Ω i −3 + Ω i +3 ) , for i = 2n + 1 (2-2) = Ω i , otherwise Where ω∈R is a variable of selecting the frequency characteristics of the coefficients where watermark is inserted. The coefficients, Ωi (for i = 2n + 1), are represented as convolution of input signal xi and a high-pass filter, hi, followed by down sampling. H( z) = Z[h i ] is written as
random noise sequence {+1, -1} are used as the watermark, by adding proportionally to the absolute of the transformed coefficients such as in [1]. The methodology for watermark detection follows basically that done in [1], with an exception of the transform.
Fig. 3. Frequency characteristics of H ( z) .
H(z) = αβω ⋅ z −5 + βω ⋅ z −4 + (αβγ + αβω + ω) ⋅ z −3 + βγ ⋅ z −2 + (α + γ + 3αβγ − 2αβω − ω) ⋅ z −1 + (1 + 2βγ − 2βω)
(3)
+ (α + γ + 3αβγ − 2αβω − ω) ⋅ z1 + βγ ⋅ z 2
Fig. 4. Detection results according to filter transform coefficient ω, where Lena image is tested and ω = 0.1 is applied in watermark insertion.
+ (αβγ+ αβω + ω) ⋅ z 3 + βω ⋅ z 4 + αβω ⋅ z 5 Since
H (e j2 πf )
f =0
= 0 , H (z) must be a high-pass filter
independently variable ω. Its frequency characteristics are shown in Fig. 3. This preserves the transparency of watermark, allowing insertion of watermark into coefficients with different frequency characteristics. In fact, Fig. 2 performs the same transform in Fig. 1. This can be easily proved. By substituting (2-1) into (2-2), (2-2) yields to (1-3). This means that the operations of [step 3] and [step 4] in Fig. 2 are equal to that of [step 3] in Fig. 1. And the other steps in two figures are just the same, respectively. As results, the modified lifting performs the same operations as usual lifting, but allows the use of various high-pass filters to select frequency characteristics of the coefficients where watermark is inserted. Now, we can insert watermark into various frequency bands using the modified lifting. The parameter ω determines the frequency bands where inserted a watermark, thus it plays an auxiliary secret key to identify the ownership. The proposed lifting based watermarking can be easily integrated in the JPEG-2000 baseline, since the insertion of watermark is performed during the operation of the lifting process. 4. EXPERIMENTAL RESULTS
The methodology for watermark insertion is as follows: First, an original image is decomposed into four multiresolution levels, where the final level is constructed by using the modified lifting and watermark is added to the intermediate transformed coefficients. Pseudo binary
Table 1. Watermark detection results in terms of correlation coefficients and PSNR in parenthesis for various attacks. Where, watermark is inserted at resolution level 4, ω= 0.1 Attack JPEG (20:1) JPEG2000 (20:1) Blurring Sharpening Rescaling
Lena
Baboon
Hat
Pepper
0.33 (32.09) 0.39 (33.85) 0.34 (32.40) 0.18 (16.37) 0.37 (30.59)
0.19 (22.72) 0.28 (24.44) 0.41 (23.28) 0.14 (11.15) 0.38 (23.21)
0.38 (31.82) 0.47 (33.73) 0.46 (33.30) 0.19 (17.30) 0.47 (31.49)
0.35 (31.55) 0.42 (33.04) 0.37 (31.12) 0.18 (16.01) 0.42 (29.13)
Table 2. Watermark detection results against various attacks in terms of correlation coefficients and PSNR in parenthesis, where Lena image is tested and ω = 0 indicates the conventional DWT based watermarking.. ω= ω= ω= ω= ω= Attack -0.2 -0.1 0 0.1 0.2 0.46 0.36 0.31 0.33 0.36 JPEG (20:1) (32.25) (32.38) (32.35) (32.38) (31.61) 0.47 JPEG2000 0.39 0.39 0.39 0.44 (20:1) (34.21) (34.38) (34.30) (33.85) (33.18) 0.41 0.36 0.34 0.39 0.48 Blurring (32.61) (32.72) (32.68) (32.40) (31.91) 0.26 0.18 0.19 0.18 0.20 Sharpening (16.37) (16.39) (16.39) (16.37) (16.35) 0.42 0.39 0.37 0.43 0.49 Rescaling (30.74) (30.81) (30.78) (30.59) (30.26)
Fig. 4 shows the detection results in terms of correlation coefficients between inserted and detected watermark sequences, according to the variable ω. In watermark insertion process, ω = 0.1 is applied. In this figure, the highest correlation coefficient occurs at the same ω as used in the watermark insertion. This demonstrates that the proposed watermarking is robust against possible alternation of inserted watermark, reinforcing security for users having auxiliary secret key, i.e. ω. Table. 1 represents correlation coefficients and PSNR (peak signal to noise ratio) with corrupted watermarked images. Where JPEG (20:1), JPEG-2000 (20:1) compression, blurring, sharpening, and rescaling are considered as attacks. It is shown that the proposed method is robust against common signal processing. Table.2 shows correlation coefficients and PSNR according to the parameter ω. When ω = 0, the proposed method can be considered as the conventional DWT based watermarking. The results demonstrate that the proposed method with different value of ω preserves the ability of watermark detection maintaining similar image quality.
5. CONCLUSIONS
A new watermarking method is proposed, which is based on the lifting of DWT. The proposed watermarking can be conveniently integrated in the JPEG-2000 baseline with a little additional computational complexity, as a watermark is inserted during the lifting process for DWT. The method can be easily applied to other kinds of filter-banks, besides 9-7 taps filter-banks used in this paper. The simulation results demonstrate that the proposed method is more secure and more robust against various attacks than usual wavelet based watermarking. 6. REFERENCES [1] Ingemar J. Cox, J. Killian, F. Thomson Leighton, and Talal Shamoon, “Secure spread spectrum watermarking for multimedia,” IEEE Trans. on Image Processing, Vol. 6, No. 12, pp.1673-1687, Dec. 1997. [2] Christine I. Podilchuk, Wenjun Zeng, "Perceptual watermarking of still images,” in Proceedings of the Workshop on Multimedia Signal Processing, Princeton, New Jersey, USA, June 1998.
[3] Young-Sik Kim, O-Hyung Kwon and Rae-Hong Park, “ Wavelet based watermarking method for digital images using the human visual system,” IEEE trans. Electronics Letters. Vol. 35. No. 6. March 1999. [4] Xiang-Gen Xia, Charles G. Boncelet, Gonzalo R. Arce, “ A Multiresolution watermark for digital images,” IEEE Int. Conf. Image Process., October 1997, Santa Barbara, CA, USA, Vol. 1, pp. 548-551. [5] D. Kundur and D. Hatzinakos, "Digital watermarking using multiresolution wavelet decomposition," in Proc. IEEE Int. Conference on Acoustics, Speech and Signal Processing, Vol. 5, pp. 2969-2972, 1998. [6] A. R. Calderbank, I. Daubechies, W. Sweldens, and B.-L. Yeo, "Wavelet transforms that map integers to integers," Applied and Computational Harmonic Analysis, Vol. 5, No. 3, pp. 332369, July 1998. [7] ISO/IEC, JPEG 2000 Verification Model 8.5 (Technical Description), 2000 [8] ISO/IEC, ISO/IEC 15444-1, Information Technology - JPEG 2000 image coding system, JPEG 2000 Final Draft International Standard, 2000 [9] G. C. Langelaar, I. Setyawn, R. L. Lagendijk, “Watermarking digital image and video data, A State-Of-The-Art Overview,” IEEE Signal Processing Magazine, Vol. 17, No. 5, pp. 20-46, September 2000.
