04273880
TRANSCRIPT
2006 IEEE International Conference on
Systems, Man, and CyberneticsOctober 8-11, 2006, Taipei, Taiwan
A Copyright Protection Scheme for Satellite Images Using SecretSharing and Wavelet Transform
Shang-Lin Hsieh and I-Ju Tsai
Abstract-This paper considers weather satellite images asthe protected host images and proposes a novel copyrightprotection scheme according to the property of the images. Thescheme contains two phases: the share image generation phaseand the watermark retrieval phase. In the generation phase, theproposed scheme first generates a cloud image from the hostimage using gray-level slicing and creates a special samplingplane image from the cloud image and the host image. Next, thescheme extracts the features from the sampling plane imageusing the discrete wavelet transform. Then, the scheme employsthe features and the watermark to generate a principal shareimage. In the retrieval phase, an expanded watermark is firstreconstructed using the features of the suspect image and theprincipal share image. Next, the scheme reduces the additionalnoise to obtain the recovered watermark, which is then verifiedagainst the original watermark to examine the copyright.
The experimental results show that the proposed scheme canresist several attacks such as JPEG compression, blurring,sharpening, scaling, and cropping.
I. INTRODUCTION
T HE rapid evolution of the Internet technology makes thetransmission of digital multimedia contents easier
nowadays than before. Copyrighted digital media such asimages, music, etc, can be copied or distributed quickly andeasily on the Internet. Media content owners are veryconcerned about the potential loss of revenue resulting fromdigital media piracy. Therefore, digital watermarkingtechniques have been proposed for copyright protection[1]-[4] or ownership identification of digital media.
Digital watermarking [4]-[8] is a technique used to protectthe digital media property. Digital watermarking embeds awatermark into the host image. Later, it extracts thewatermark from the suspected image for verification.A watermark can be embedded in the spatial domain or the
frequency domain. In the spatial domain, the watermark isembedded by changing the pixel values directly. In thefrequency domain, the pixel values are transformed tofrequency coefficients and the watermark is then embeddedby modifying the coefficients. Generally, the robustness ofthe watermark embedded in the frequency domain is oftenstronger than that in the spatial domain.
The problem with both of the above watermarkingtechniques is that they modify some pixels of the host imagedirectly or indirectly, which decreases the image quality.Moreover, they are not suitable for unchangeable images such
Manuscript received March 15, 2006.The authors are with the Department of Computer Science and
Engineering, Tatung University, Taipei, Taiwan (corresponding e-mail:[email protected]).
as satellite images or medical images.To preserve the image quality, some researchers looked for
the techniques that do not alter the original image. Chang andChuang [2] applied the visual secret sharing scheme (VSS) onthe watermark to generate two share images and store one ofthem for later verification. In the verification phase, theirscheme generates the other share image and superimposes thestored one on the generated one to retrieve the watermark.The problem with the scheme, however, is that itsperformance deteriorates as the JPEG compression ratioincreases.The author once proposed a scheme [1] to protect copyright
for gray-level images using secret sharing. However, thescheme did not make any assumption about the host image. Ifwe know in advance the kind of host images and design ascheme that takes advantage of the information, better resultcan be achieved. This paper proposes a new copyrightprotection scheme that takes weather satellite images as thehost images. The proposed scheme generates a cloud imagefrom the host image and then uses a specific sampling methodto generate a sampling plane image from the cloud image andthe host image. The sampling plane image is then used forfeature extraction.The proposed scheme has the following advantages. First, it
does not modify the host image. Therefore, it is suitable forthe application in which the modification of the image is notallowed. For example, satellite images cannot allow anymodification on them because that will affect their precisions.Second, the scheme is secure. By applying the technique ofsecret sharing, only the user who has the share image canretrieve the watermark. Last, the scheme is robust. Theexperimental results show the scheme can resist five kinds ofcommon image processing operations, including JPEGcompression, bluffing, sharpening, scaling, and cropping.
The organization of the paper is as follows. Section IIdescribes the related background of the proposed scheme.Section III explains the proposed sampling method. SectionIV details the proposed copyright protection scheme. SectionV examines the experimental results. Finally, section VIdraws the conclusion.
II. RELATED BACKGROUND
The proposed scheme uses the discrete wavelet transform(DWT), secret sharing, and Gray-level slicing. This sectionbriefly explains the above techniques.
A. Discrete Wavelet TransformThe basic idea of the DWT for an image is described as
1-4244-0100-3/06/$20.00 C2006 IEEE 501
follows. An image is first decomposed into four sub-bandsLL1, LHl, HLl, and HH11. The LL1 sub-band can be furtherdecomposed into four sub-bands LL2, LH2, HL2, and HH2.The same decomposition procedure can be applied until thereis only one coefficient in the LL sub-band. Fig. 1 shows theimage "Lena" and the result after two-level DWTdecomposition.
_ _sJI {AX s
U~~ad/
l/f£ f /f/f
Fig. 1. Two-level DWT decomposition ofLena
B. Secret SharingThe proposed scheme applies the secret sharing scheme
[11] [12] to protect image copyright. The secret sharingscheme splits an image into n different shares. The image canbe retrieved with more than k (k < n) shares. The size of theretrieved image is expanded because each pixel is mappedinto a block consisting of several sub-pixels. The effect iscalled pixel-expansion. Fig. 2 shows an example of the secretsharing scheme. Fig. 2 (a) is the original image. Fig. 2 (b) and(c) are the share images produced from the original one. Fig. 2(d) is the retrieved image, which contains the information thatcan be recognized visually with the human eyes. Note, theretrieved image is four times larger than the original one andcontains noise because ofpixel-expansion.
_ay II
10) (:Ct:) (d)Fig. 2. An example of the secret sharing scheme; (a) the originalimage; (b), (c) the share images; (d) the retrieved image
C. Gray-level SlicingGray-level slicing can highlight a specific range of gray
levels in an image. Normally, the gray level of the cloud ishigh and therefore the scheme keeps the value between 127and 255 to generate the cloud image. Fig. 3 shows thetransformation.
2555
127 ,
0 127 255Fig. 3. The transformation of the gray-level
III. THE PROPOSED SAMPLING METHOD
The scheme uses weather satellite images as the host image.It applies gray-level slicing to the image, and then generates anew sampling plane image based on the proposed samplingmethod. This section explains the method.
A. The Rule ofthe Proposed Sampling MethodAt first, a cloud image B is generated from the host image
H. Then, the new sampling plane image F is generatedaccording to (1).
F(x, y) = fH(x,y) ifxmod(8) = 0,1, 2,3 (1){B(x, y) ifx mod(8) = 4,5,6,7
where x andy are the coordinates of the image. Fig. 4 (a) and(b) are the weather satellite image and the cloud imagerespectively. Fig. 4 (c) is the new sampling plane image.
(a) (b) (c)Fig. 4. The proposed sampling method; (a) the host image; (b) the cloudimage; (c) the sampling plane image
B. The Consideration ofthe Proposed Sampling MethodThe arrangement of the proposed method can be described
from another point of view as follows. The sampling planeimage is obtained by combining half of the original image andhalf of the cloud image, which are interlaced together.
The benefit of the arrangement is that many attacks will beoffset by the gray-level slicing while the original features canstill be kept because of the half of the original image.
IV. THE PROPOSED COPYRIGHT PROTECTION SCHEME
The proposed scheme contains two phases: share imagegeneration phase and watermark retrieval phase. Fig. 5shows the block diagram of the proposed scheme. Thefeatureextraction in the share image generation phase first generatesa cloud image from the host image. Then, it creates asampling plane image using (1) and extracts the features fromthe sampling plane image. Afterward, the watermarkscrambling disarranges the watermark with a secret key.Finally, the encoding uses the features and the scrambledwatermark to generate the principal share image, which issaved and will be used for watermark retrieval To retrievethe watermark, the feature extraction in the watermarkretrieval phase first generates the cloud image from thesuspect image. Then, it creates a sampling plane image andextracts the features. The decoding next
502
Sampling l - , ) _ )plane image_ _
Scrambled Watermarkll ~~~~~~~~~~Watermark Reduction3 >' f~~~~~~~~~~~~~~Watermark Extnde
, t~~~~~~~~~~Unscrambling0 Watermark
--Share Image Generation Phase Watermark Retrieval Phase ~
Fig. 5. The block diagTam ofthe proposed scheme
uses the features and the previously saved principal shareimage to retrieve the scrambled watermark. Afterward, thewatermark unscrambling rearranges the scrambledwatermark. Finally, the watermark is reduced to obtain therecovered watermark. The recovered watermark is then usedto verify the copyright.
A. The Major Parts in the Proposed SchemeThere are four major parts in the scheme. The following
subsections describe the details of the four parts.1) Scrambling and Unscrambling Watermark: The
proposed scheme uses Torus-automorphism [13] [14] toscramble the watermark in the share image generation phaseand to unscramble the scrambled watermark in the watermarkretrieval phase. Fig. 6 shows an example of a scrambledwatermark after five iterations.
Fig. 6. The scrambled watermark
2) Feature Extraction. Before feature extraction, Thecloud image must be generated from the host and suspectimages. Next, a sampling plane image is generated from thehost image and the cloud image by the method described inSection III. Then, the sampling plane image is divided intonon-overlapping blocks of size 8 x 8 and the elements of eachblock are then transformed to DWT coefficients.
After applying two-level DWT, there are four coefficientsin the LL2 sub-band of each block. Let M be the average ofthe four coefficients. Then, M should satisfy one of thefollowing conditions:
1. Only one coefficient is smaller than M.2. Only two coefficients are smaller than M3. Only one coefficient is greater than M.4. All of the four coefficients are the same and therefore
equal to M.A feature type is then obtained from the above rule. Let n
be the feature type, which represents the number ofthe above
conditions, thent1, ifM satisfies condition (1)2, ifM satisfies condition (2)
n=n3, ifM satisfies condition (3)
4, ifM satisfies condition (4)
(2)
3) The Encoding and Decoding: According to the secretsharing scheme, two share blocks are created: principal shareand complementary share blocks. Each of the two shareblocks contains 2 x 2 pixels. Each pixel of the watermark canbe mapped into a share block of 2 x 2 pixels according to thepixel value of the watermark. The share blocks form the shareimage.
Table I lists the mapping table used in the proposed scheme.A pixel ofthe watermark will be mapped into a block of size 2x 2, which is called pixel-expansion.The encoding process uses the extracted feature types and
the watermark to generate the share blocks of the principalshare image. The principal share blocks are generatedaccording to the description in section IV.B.The decoding process applies XOR operation (Table II) on
the complementary share blocks obtained from the suspectimage and the corresponding share blocks of the previouslysaved principal share image (generated from the original hostimage) to retrieve the scrambled watermark. Then, thescrambled watermark is unscrambled to obtain the visuallyrecognizable watermark.
4) Watermark Reduction: When the original watermarkpixel is white, some redundant noise on the background oftheunscrambled watermark will be generated due to thepixel-expansion effect of secret sharing. To regain theoriginal watermark from the unscrambled watermark, awatermark reduction process is used to remove the redundantnoise. According to the generated block by XOR, one of thefollowing two steps will be adopted:
1. when the number of white pixels of the generated blockis equal to 1 or 0, the block is reduced to a black pixel.
2. when the number of white pixels of the generated blockis greater than 1, the block is reduced to a white pixel.
503
Table III lists the reduction conditions and thecorresponding actions. Fig. 7 shows an example of a
recovered watermark after reduction.TABLE I
THE MAPPING TABLE FOR ENCODING AND DECODING
P-Share: PTiLeipal chase C-Share: Caoiplenceiar shae
TABLE II
THE XOR RULE
Pixel-I Pixel-2 Pixel-IXOR Pixel-2
0 0
TABLE IIITHE REDUCTION CONDITION AND THE CORRESPONDING ACTIONCondition Action Condition Action
Reduce the result block Reduce the result block
into a white pixel into a block pixel
The number of The number ofwhite pixels of Example white pixels of Example
the block is hie block is
greater then I EEF'{ equal to 1 or 0
Fig. 7. The example of the watermark reduction
B. The Procedure ofShare Image GenerationThe share image generation phase first generates a cloud
image from the host image. Next, it creates a sampling planeimage, from which it extracts the features. Then, it scramblesthe watermark with a secret key. Finally, it generates theprincipal share image using the features and the scrambledwatermark. The following lists the detailed steps.
Share Image Generation Procedure
Input: A host image H(NxN), a watermark W(N!8xN/8), and a
secret key for scrambling.Output: The principal share image S(N/4xN/4) used toretrieve the watermark.Step 1 Use Torus-automorphism and the secret key to
scramble the watermark W into W.Step 2 Generate the cloud image B from the host image H.Step 3 Sample the host image H and the cloud image
B to from the sampling plane image F according to (1).Step 4 Divide the sampling plane image F into
non-overlapping blocks of size 8 x 8 F(k), kl1, 2,N18 x N18, and divide the scrambled watermark W intoa binary digits sequence W (k), kl=1,2,..., N18 x N18.
Step 5 Let k be 1.
Step 5.1 Transform the sampling plane block F(k) intofour coefficients by two-level DWTdecomposition.
Step 5.2 Calculate the average M of the fourcoefficients.
Step 5.3 Obtain the feature type n according to (2).Step 5.4 Based on Table I, use the feature type n, the
location of average M of the four coefficients,and the pixels of the corresponding scrambledwatermark block W(k) to generate the principalshare block S(k).
Step 5.5 Increase k by one. If k < N18 x N18 go to Step5.1.
Step 6 Output the principal share image S consisting of theshare blocks.
The generated principal share image S is then saved andwill be used in the watermark retrieval phase.
C. The Procedure of Watermark RetrievalThe watermark retrieval phase first generates a cloud
image from the suspect image. Next, it creates a samplingplane image, from which it extracts the features. Then, it usesthe features and the previously saved principal share image toretrieve the scrambled watermark. Finally, it rearranges the
504
Fealse Mean Vahle The watnecaik P-Share Thewatreaias P-Share
Type n Locatior pixel is whicte XOR- pixel is black XOR
C-Share C-ShareP-Share C-Share P-Share C-Slhare
1 acicbc,d
b<M< %C,& 1 3
ccl,d1 EM 'N
3
Hi 0c<M<da ,e
a,c<Mea1b,cd
2 avc<Wch7d 6 EEO E 0
bd.M a,c , ] C NE%b,cd.-M<a E J k d J 0a,c,d.cMebe L A AEa,b,d.MecM ~
a,b,c<M<d
4 M=ab=c=dC*
_v Th L7ElilV IO{Erflk. T.T iEEEi-.nl tW inialb I ne mm ooetirierits c(tuteLLz Rmoamka isaaimmopiei., c- ismtmmpil:..!-L-
!"gio. ri&, c is at tle bottam lefl, and d is al the boffornri& pc)sftic)rL)
scrambled watermark and reduces the watermark to obtainthe recovered watermark. The following lists the detailedsteps.
Watermark Retrieval Procedure
Input: A suspect image H'(N x N), the principal share imageS(N/4xN/4), and the secret key for unscrambling.Output: The recovered watermark WR(N/8xN!8)Step 1 Generate the cloud image B' from the host image H'.Step 2 Sample the host image H' and the cloud image
B' to generate the sampling plane image F' according to(1).
Step 3 Divide the sampling plane image F' intonon-overlapping blocks of size 8 x 8 F'(k), k=1, 2, ....N18 x N18, and divide the principal share image S intonon-overlapping blocks of size 2 x 2 S(k), k= 1,2,..., N18x N18.
Step 4 Let k be 1.Step 4.1 Transform the sampling plane block F'(k) into
four coefficients by two-level DWTdecomposition.
Step 4.2 Calculate the average M of the fourcoefficients.
Step 4.3 Obtain feature type n according to (2).Step 4.4 Based on Table I, use the feature type n, and the
location of average M of the four coefficients togenerate the complementary share block S'(k).
Step 4.5 Apply XOR operation on the complementaryshare block S'(k) and the principal share blockS(k) to produce the corresponding scrambledwatermark W'(k).
Step 4.6 Increase k by one. If k < N/8 x N18 go to Step4.1.
Step 5 Use Torus-automorphism and the secret key tounscramble the watermark W' to '.
Step 6 Use Watermark reduction process described inSection IV.A.4 to reduce the unscrambled watermarkW' and remove the noise to obtain the recoveredwatermark WR.
V. EXPERIMENTS
We conducted some experiments to evaluate the feasibilityof the proposed scheme. Fig. 8 (a) shows the test weathersatellite host image. Fig. 8 (b) and (c) shows the watermarkand the generated principal share image. We used acommercial image processing software, Ulead PhotoImpact8.0, to simulate different kind of attacks, including JEPGcompression, blurring, sharpening, scaling, and cropping.
(a) k(,Fig. 8. (a) The test weather satellite image (512x512); (b) Thewatermark (64x64); (c) The principal share image (128x 128)
The Accuracy Rate (AR) is used to measure the differencebetween the original watermark and the recovered one. AR isdefined as follows:
AR - CP (3)NP
where NP is the number of pixels in the original watermarkand CP is the number of correct pixels obtained by comparingthe pixels of the original watermark with the correspondingones of the recovered watermark. The more closely ARapproaches one, the more closely the recovered watermarkresembles the original one.
Table IV lists the experimental results. On the left of thetable are the results generated from the proposed schemewhile on the right are the ones generated using the author'sprevious scheme for gray-level images [1], which extracts thefeatures only from the host image and does not take theadvantage of the cloud image.
TABLE IVEXPERIMENTAL RESULTS OF THE PROPOSED AND PREVIOUS SCHEMES
AR.............G.....
Blurring
AR
AR
:ahugX
AR
Cropped area
AR
25%.....o 9..0 95" 909
50%0 9651
75%
0.9744
''..A .'3 2
............. .................. ... ...........
0.8994 0.9126 0.9375. . .... ............. ...... .......... . .
3 2
0.8394 0 8923 0.9504
64x64 128x128 256x256
0.8000 0.8845 0 9551
50% 25% 10%.......... ...
0.8777 0.9380 0.9763kik..
25%O 50% 75%0 9221 0 9487 0 9626
3 2 1
0 8848 0 9038 0 9329.................. ....... ...... ........
3 2 1
0 8472 0.8928 0.9426
64x64 128x128 256x256
0 7183 0.8567 0.9456
. .................. . ..........50% 25% 10%
0.8596 0.9353 0 .9729
505
The proposed schem0e fI The previqou sce:me:#::. -;:
1-
The experiment shows the recovered watermnarks can all beeasily recognized after the attacks. It also shows the proposedscheme almost outperforms the previous one for each attackexcept for two items in sharpening. The difference however isquite insignificant (less than 0.008).
In addition, a uniqueness experiment was conducted tocheck whether a different image would be mistaken for theoriginal one. The features of various images should bedifferent, so the recovered watermark from an image differentfrom the original one should contain nothing more than noise.
Fig. 9 shows another test weather satellite image and therecovered watermark, which is different from the originalimage. The accuracy rate of the test image is 0.6672. Therecovered watermark contains no information about theoriginal watermark.
Fig. 9. The test weather satellite imagewatermark
and the recovered
[7] Narges Ahmidi and Reza Safabakhsh, "A Novel DCT-based Approachfor Secure Color Image Watermark," Proceedings of the InternationalConference on Information Technology, 2004.
[8] Chun-Hsien Chou and Tung-Lin Wu, "Embedding Color Watermarksin color Images,' IEEE, 2004
[9] D.J.Fleet and D.J.Heeger, "Embedding Invisible Information in ColorImage," in proc.ICIP'97, pp.532-535, Oct. 1997.
[10] M. kutter, F.Jordan and F. Bossen, "Digital Signature of Color ImagesUsing Amplitude Modulation, " in Storage and Retrieval for Image andVideo Database V, SPlE, vol.3022.pp. 518-526,San Jose, Ca Febuary8-14, 1997.
[11] Lukac, R.; Plataniotis, K.N., "A secret sharing scheme for imageencryption", 46th International Symposium, pp. 549- 554, June 2004.
[12] M. Naor and A. Shamir, "Visual cryptography", Advances inCryptology-EUROCRYPT'94, Lect. Notes in Comput. Sci. vol. 950,pp.1-12, 1995.
[13] G. Voyatzis and I. Pitas, "Applications of torus automorphisms inimage watermarking," in Proc. Int. Conf Image Processing (ICIP), vol.3, Lausanne, Switzerland, Sept. 1996, pp. 237-240.
[14] Chin-Chen Chang, Ju-Yuan Hsiao and Chi-Lung Chiang, "An ImageCopyright Protection Scheme Based on Torus Automorphism", FirstInternational Symposium on Cyber Worlds (CW'02) November 06 - 08,2002
VI. CONCLUSIONS
In this paper, we proposed a copyright protection schemefor weather satellite images using secret sharing and thediscrete wavelet transform (DWT). The scheme outperformsthe previously proposed scheme while still preserving thefollowing advantages:
(1) it does not modify the host image, and therefore issuitable for unchangeable weather satellite images,
(2) it is secure because ofthe employment of secret sharingand Torus-automorphism transformation, and
(3) it is robust according to the experimental results, whichshows the accuracy rates are almost higher than 85%.
REFERENCES
[1] Shang-Lin Hsieh and Bin-Yuan Huang, "A Copyright ProtectionScheme for Gray-Level Images Based on Image Secret Sharing andWavelet Transformation," Proceedings of International ComputerSymposium, 2004.
[2] C.C. Chang and J. C. Chuang, "An image intellectual propertyprotection scheme for gray-level images using visual secret sharingstrategy," Pattern Reconition Letters, vol. 23, pp. 931-941, June 2002.
[3] Ren-Junn Hwang, "A Digital Image Copyright Protection SchemeBased on Visual Cryptography," Tamkang Journal of Science andEngineering, vol.3, No. 2, pp. 97-106,2000.
[4] San-Hwa Wu and Hsiu-Feng Lin, "A Non-Embedded WatermarkingScheme for Non-Distortion Digital Product Copyright Verification,"2004
[5] Xiang Zhou, Xiaohui Duan and Daoxian Wang, "A Semi-FragileWatermark for Image Authentication," Proceedings of the 10thInternational Multimedia Modelling Conference, 2004.
[6] Xiaoqiang Li and Xiangyang Xue, "Improved Robust Watermarking inDCT Domain for Color Image," Proceedings of the 12th InternationalConference on Advanced Information Networking and Network, 2004.
506