Ablind Associative Watermark Detection Scheme Using Self-Embedding Technique

ABlind Associative Watermark Detection Scheme Using Self-embedding Technique

Chin-FengLee and Huai-En Lee

Department of Information Management,

ChaoyangUniversity of Technology,

168Jifong E. Rd., WufongTownship,TaichungCounty 41349, Taiwan, R.O.C.

E-mail: {lcf, s9514614}@cyut.edu.tw

Responsible for correspondence:

Dr. Chin-Feng Lee

Department of Information Management,ChaoyangUniversity of Technology,

168, Jifong E.Rd.,WufongTownship, Taichung41349, Taiwan

E-mail:

TEL: 886-4-23323000-4293

FAX: 886-4-23742337

ABlind Associative Watermark Detection Scheme Using Self-embedding Technique

Abstract

This paper proposed a blind associative watermarking by means of self-embedding technique to improve the ability of detecting watermarks proposed by Shen et al.[4]. From the experimental results, the proposed method ensures the convincing robustness without the assistance of original host image under around 99% confidence when identifying if a given detection image has the watermark embedded within even though it has met with some image processing.

1. Introduction

Due to the development of the Internet, multimedia filescan nowbe shared quickly. In addition, multimedia images can be easily modified, causing the owner possibly to lose their copyright when such images are transmitted and shared over anopen channel. Copyright protection that does not degrade image quality is becoming an important issue.

Many watermarking techniques[2], [5], [6], [7]have beenproposed to solve the copyright protection problem for multimedia images.Robustness and image quality are two crucial issues for the watermarking techniques.Traditional watermarking techniquesgenerally maintain the original image to provide information for the watermarking detection scheme. Thus, it is a waste of time and space to search for and keep the original image. In order of solve this problem, a blind watermarking technique[3]is developed that does not require to keep the original image but can still process the detection of watermark. In 2005, Shen et al.[4] proposed a newwatermarking detection technique based on association rule mining. Their experimental results also showed that association rule mining can make the watermarking technique more robust and provide ahigher quality for watermarked image. However, Shen et al.’s watermarking detection scheme is not able to provide the property of blindness. That should be consuming a large amount of resources to detect the watermark.

This paper proposed a blind associative watermarking by applying the self-embedding technique to embed the rule information associated to the original image into those of the watermarked image. This paper is organized as follows. In Section 2, we review the association rule mining and associative watermarkingscheme proposed by Shen et al. In Section 3, the details of our proposed scheme are presented. In Section 4, the experimental results are discussed. Finally, we make conclusions in Section 5.

2. Related Works

This section provides a comprehensive review of previous works in association rule mining and the associative watermarking detection scheme as well.

2.1 Association rules

Data miningfinds useful information among a large amount of data. The association rulemethod was proposed in 1993 [1]can find the state and frequency of data as well as the relationship among data.

Association rules are generally applied into a transaction database. Each of transaction consists of several items which come from a set of items as labeled as It. Using the association rules algorithm, we can define the association rule as follows. Suppose X and Y are itemsets, where XIt, YIt, X∩Y=, and |X|+|Y|=K. Then, an association rule is defined as R: X→Y[s, c] upon the K-itemset, s and c stand for the support and confidence values of R, respectively. In addition, the association rule should satisfy the requirement of that both the support and confidence values are greater than the predetermined minimum support and confidence values.

2.2Associative watermarking technique

In 2005, Shen et al. [4] proposed a new watermarking detection scheme called associative watermarking. They used association rules to find the relationship between the original image and the watermark image, embedding the relation rules into the watermarked image so that the association rules can beretrieved to be verified whether the watermarked image suffered an attack.

2.2.1 Association image rules and the related notation description. A digital image is composed of a red layer, a green layer, and a blue layer. Generally speaking, the human visual system (HVS) is sensitive to the green layer and insensitive to the blue layer. Therefore, for each block b of image I, we use Equation (1) to obtain B1I(b) for the green layer and Equation (2) to obtain B2I(b)for the blue layer.

, / (1)
. / (2)

Here, B1I(b) stands for the first block’s item value for the association rule as introduced as below for every block b, n is the number of neighboring blocks, IG(b) is the average number of blocks, and IG(b’) is the average number of the neighboring blocks. Subsequently, the second block’s item value for the association rule follows Equation (2) for every block b, where DCT(IB(b)) is the DCT transformation of block b and L is the number of coefficients with low frequency. For each block b, we use Equations(1) and (2) to obtain the association rule set {B1I(b)→B2I(b)} of image I. We use a similar calculation for the watermarked image W and the detection image D to obtain the corresponding association rule sets{B1W(b)→B2W(b)} of W and {B1D(b)→B2D(b)} of D, respectively. Then, Equations (3) and(4) are applied to restrict the values of B1(b) and B2(b) for eachblock bwithin the range of [0,Ω-1].

, / (3)
, / (4)
whereΩ and δ are within the range from 1 to 256 for image I.

Finally, the set of original image rules is defined as RI:{Q1(B1I(b))→Q2(B2I(b))}; the set of watermark image rules is defined as RW :{Q1(B1W(b))→Q2(B2W(b))}; and the set of detection image rules is defined as RD :{Q1(B1D(b))→Q2(B2D(b))}, respectively.

2.2.2 Embedding associative watermarks.The associative watermark embedding procedure can be divided into two phases: obtaining the image rules and obtaining the alignment rules. In first phase, both the original image and the watermarked image are divided into image blocks in order to obtain association rules RI and RW. In the second phase, the first step is to mineRW by maximum confidence in order to obtainRM. Then, weobtainall rulesRA with the same B1(b)values forRI and RM. For each rule RA,we update thesecond item values such that Q2(B2I(b))=Q2(B2W(b)). Finally, for each rule RA,we find alignment with the watermark rule;that is, we achieve the goal of the watermark rule hiding.

2.2.3 Detection of associative watermarking.The watermarked image can be calculated with a similar degree of association rules via the detection procedure. Associative watermarking detection can be divided into two phases.In first phase, the original image I, the detection image D, and the watermarked image W are obtained by finding the values ofRI, RD, and RW, like in the first phase of the watermark embedding procedure. In the second phase, the value of RMis obtained by mining maximum confidence with RW and calculating the frequency of each rule i from RM in RI and RD and defined as RI,Count(i) and RD,Count(i).

, / (5)

The similarity S as shown in Equation (5) is proposed to evaluate how similar it is between the original image and the detected image,where N is the number of association rules of RM and Ci is the first item value of the association rules I in RM.Finally, for apredeterminedthreshold T, S≧Tindicates that the detection image D hides the associative watermark rules; the detection image suffers from some intentional attacks, otherwise. The experimental results show that thescheme developed byShen et al.[4] can resist general image-processing attacks and that the accuracy rate is greater than 90% as well. That means that Shen et al.’s proposed associative watermarking is robust.

3.Proposed Blind Associative Watermark-ing

The digital watermarking technique has been extensively applied to image protection. The associative watermarking technique proposed by Shen et al. obtains high robustness but cannotachieve blindness, because it needs the original image to obtain the detection rules. Maintaining the original image will cause problemswith storage and searching time.

For this reason, we propose a blind associative watermarking scheme by applying self-image technique to embed the association rules related to the original image into the middle-frequency area of the DCT transformed watermarked image In this manner, the proposed can preserve robustness as well. In this section, we present our proposed method, which consists of embedding and detection procedures.

3.1 The embedding procedure of blindassociative watermarking

In order to achievethe requirement of blind watermarking, we adjust the association rules of the original image such that the rules representing the watermark image can be embedded. This procedure is divided into the following three steps as shown in Figure 1.

Step 1

In this step, the original image I and the watermark image Ware divided into m by m image blocks b.

Figure 1. The embedding procedure of blind associative watermarking

For each block b,we obtain the green and blue layers and then extract B1I(b) and B2I(b) by Equation (1) and (2), respectively. Next,Equation (3) and (4) are applied to obtain the association rule sets RI: Q1(B1I(b))→Q2(B2I(b)) of I and RW: Q1(B1W(b))→Q2(B2W(b)) of W, respectively.

Step 2

In this step, the second item values of RIareadjusted. The adjustment way is described as follows.

Firstly partition RW into K set of association rules according to the first item values of RW. For each rule set, pick up that with maximum confidence value to generate the set RM.

Secondly, for each rule r whose first item value is the same as that of the rule r’ in RM, we adjust the second item value of r to be that of rule r’ such that we can obtain RI’:{Q1(B1I(b))→Q2(B2M(b))}.

Last, we reset LDCT-coefficients of block b to be the value of B2M(b) such that the association rule corresponding to the watermark image can be embedded.

Step 3

In order to achieve blindness, we embed theadjusted rulesof block b into a corresponding masking block ofthe watermarked image to allow for detection when the original image is not available.For robustness consideration, the masking block can be determined block b. The reason is if we take other L DCT-coefficients next to those where the adjusted rules were embedded within to be the locations of rule self-embedding, then the cascading destroyed will be reduced to the minimum even when the block b is lost or damaged. The self-embedded value for each DCT-coefficient is still B2M(b). Finally, all the second item values corresponding to the blue layer have to be transformed into those in the spatialdomain by IDCT in order to produce the watermarked image.

3.2 The detection procedure of blind associative watermarking

The detection procedure is divided into two steps, as shown in Figure 2. The first step is to extract the association rules of detection image D and of the watermark image W. The second step is to calculate the similarity value of S and determine if the detection image has ever been attacked by comparing with a given threshold T.

Step 1

First, the detection image D and the watermark image Ware divided into m by m non-overlapping image blocks. For each block b, like inStep 1 of the embedding procedure, we extract the values of B1D(b)) and B2D(b) by Equation (1) and (2)of D.

Figure2. The detection procedure of blindassociative watermarking

Next, we quantize Q1(B1D(b)) and Q2(B2D(b)) by Equation (3) and (4), respectively. Each of the Llow DCT-coefficientsin the upper left corner of block b is denoted as Radjust(b),and each of the next L DCT-coefficients then the rule set is denoted asRembed(b). Moreover, the association rulesRW of watermark image have to be reproduced again.

Step 2

In this step, we first generateRM by mining maximum confidence with RW and then calculate the confidence of each rule within Radjust and Rembed.We set a threshold of confidence TH. For each block b, the evaluation of similarity s(b), for those rules belonging to RM, between two setRadjust(b)and Rembed(b) can be considered in the following three cases.

Case 1: The second item values arethe same, and theconfidencedifference is less thanTH.

This is the best condition. Thisblock similarity s(b) is set to beone plus the confidence average of Radjust(b) and Rembed(b).

Case 2: The second item values arethe same, and theconfidence difference is greater thanTH.

The block b of detection image may have beenattacked in this condition. Therefore, This block similarity s(b) is set to be one minus the absolute of confidence difference betweenRadjust(b) and Rembed(b).

Case 3: The second item values are different.

The block b ofthe detection image may be nonexistent. So, thisblock similarity s(b) isset to be one minus the absolute of confidence difference betweenRadjust(b) and Rembed(b), and further then divided by the double ofdifference between thesecond item values.

The similarity S of the detection and watermark images can be regarded as the average of all the values of similaritys for all blocks. Finally, a threshold T is predetermined to judge whether the set of associative watermark rules for detection image D exist. When S≧T, the detection image D is considered as that thewatermark exists, otherwise, the watermark might be destroyed or not exist within the detection image D.

4. Experimental Results

In our experiments, the original image is a 512× 512 color image of Lena as shown inFigure 3(a), and the watermark image is a 128× 128 color emblem of the Chaoyang University of Technology of Taiwan as shown inFigure 3(b).

Figure 3. Image presentation for experiments (a) Original image Lena; (b) Watermark image logo of CYUT, Taiwan

There are two experiments conducted for the purposes of finding an appropriatethreshold valueT to identify the difference between the detection and original images, and for verifying the ability of robustness of the proposed blind watermarking scheme.

In Table 1, the experimental results show the proper threshold T can be set as 1.5 withthe confidence of around 99%

Table 1. The Similarity values of 1,000 color
non-watermarked images
Similarity / 1.2 / 1.3 / 1.4 / 1.5 / 1.6 / 1.7 / 1.8 / 1.9
Number ofimages are identifiedout / 876 / 896 / 917 / 990 / 990 / 998 / 999 / 1000

For the feasibility tests, we use Lena as the detection image. First, the threshold of confidence TH, the quantization parameters Ωand δ according to Equation (3) and (4) are set as 0.01, 8, and 30, respectively. After applying the embedding procedure of the proposed blind associative watermarking scheme,Figure 4 shows the watermarked image of Lena whose peak signal-to-noise ratio (PSNR) is 38.42.

Figure 4. Watermarked image of Lena
Table 2. Experimental results of Lena image.
Image processing type / Similarity value / Using T= 1.5 / PSNR (dB)
No-embedded / 0.8804 / No / -
Attack-Free / 2 / Yes / 38.42
Blurring / 1.9909 / Yes / 37.60
More blurring / 1.904 / Yes / 35.68
Sharpening / 1.907 / Yes / 34.33
More sharpening / 1.1543 / No / 29.32
Histogram / 1.8119 / Yes / 25.21
Brightness adjustment
(+40, +50, +60) / 1.9957, 1.9883, 1.9673 / Yes, Yes, Yes / 21.01, 19.22, 17.79
Gaussian noise
(10%, 20%, 30%) / 1.6319, 1.4599, 0.9383 / Yes, No, No / 24.49, 19.19, 16.57
JPEG compress
(80%, 60%, 40%) / 1.9931, 1.9919, 1.3943 / Yes, Yes, No / 36.89, 36.03, 36.30

Photoshop8 is used to perform theblurring, sharpening, histogram, etc, as shown in Table 2, such that the proposed scheme can tell if the detection images have the watermark embedded within even they suffer from some degree of image processing or modification.

5.Conclusion

Inspired by the robustness associative watermarking [4],a novel watermark detection is proposed in a blind manner by using the technique of self-embedding. The proposed scheme can hide the adjustedassociation rules into the watermarked image and the rulescan be used to assist be detecting whether a watermark is embedded or not in the future. Form the experimental results, our scheme has the degree of confidence of around 99% to identify if a given detection image has the watermark embedded within even though it has met with some image processing or modification.

Reference

[1]R.Agrawal, T.Imielinski, and A. Swami, “Mining Association Rules between Sets of Items in Large Databases,” Proceedings of the 1993 ACM SIGMOD international conference on Management of data, 1993, pp. 207-216.

[2]L.H. Chen and J.J. Lin, “Mean Quantization based Image Watermarking,” Image and Vision Computing, 2003, Vol. 21, pp. 717-727.

[3]I. Hong, I. Kim and S.S. Han, “A blind watermarking technique using wavelet transform,”Proceedings of IEEE International Symposium, 2001, Vol. 3, pp. 1946-1950.

[4]J.J. Shen, and P.W. Hsu,“Embedding Watermark with Association Rules Alignment,”Lecture Notes In Computer Science,2005,pp. 1159-1167.

[5]M.S. Wang, C.C. Chang, K.F. Hwang and J.S. Pan, “An Embedding Algorithm for Multiple Watermarks,” Journal of Information Science and Engineering, 2003, Vol. 19, pp. 381-395.

[6]Y.T. Wu and F.Y. Shin, “Digital Watermarking Based on Chaotic Map and Reference Register,” Pattern Recognition, 2007, Vol. 40, No. 12, pp. 3753-3763.

[7]Y. Xin, S. Liao andM. Pawlak, “Circularly Orthogonal Moments for Geometrically Robust Image Watermarking,” Pattern Recognition, 2007, Vol. 40, No. 12, pp. 3740-3752.