Joint robustness and security enhancement for feature-based image watermarking using invariant feature regions

(1)

Joint robustness and security enhancement for feature-based image

watermarking using invariant feature regions

Jen-Sheng Tsai

a

, Win-Bin Huang

a

, Yau-Hwang Kuo

a,n

, Mong-Fong Horng

b

a

Center for Research of E-life Digital Technology, Department of Computer Science and Information Engineering, National Cheng Kung University, Tainan 701, Taiwan

b

Department of Electronic Engineering, National Kaohsiung University of Applied Sciences, Taiwan

a r t i c l e

i n f o

Article history: Received 16 March 2011 Received in revised form 28 November 2011 Accepted 29 November 2011 Available online 8 December 2011 Keywords:

Digital image watermarking Feature detector

Knapsack problem Genetic algorithm Differential entropy

a b s t r a c t

Local image features have been widely applied in feature-based watermarking schemes. The feature invariance is exploited to achieve robustness against attacks, but the leakage of information about hidden watermarks from publicly known locations and sizes of features are often unconsidered in security. This paper, therefore, proposes a novel image watermarking approach, which adopts invariant feature regions to jointly enhance its robustness and security. Initially, circular feature regions are determined by the scale-adapted auto-correlation matrix and the Laplacian-of-Gaussian operation. Leakage of secret information is also controlled carefully during feature detection procedure. An optimal selection process formulated as a multidimensional knapsack problem is then proposed to select robust non-overlapping regions from those circular feature regions to resist various attacks. This process is implemented by a genetic algorithm-based approach, and incorporates randomization to mitigate the security risk. Finally, each selected region is normalized to obtain a geometrically invariant feature region, and embedded with a region-dependent watermark to overcome the weakness of multiple-redundant watermarks. The evaluation results based on the StirMark benchmark present the proposed scheme can tolerate various attacks, including noise-like signal processing and geometric distortions. A security analysis in terms of differential entropy also conﬁrms the security improvement of the proposed method.

1. Introduction

Digital watermarking, which is regarded as a useful approach for copyright protection, content authentication, and transaction tracking, has been widely applied to image, audio, and video. In all of these applications, the effectiveness of a digital watermarking algorithm depends

on its ability to resist various attacks. According to different intentions of attacks, there are two criteria, robustness and security, which should be considered in the design of digital watermarking schemes [1–7]. Robustness deals with blind attacks that try to destroy or invalidate hidden watermarks without exploiting knowledge of the watermarking algorithm. The robust-ness measurement for watermarking schemes is to eval-uate their ability to successfully detect the hidden watermark after blind attacks. In general, robust water-marking schemes are developed to resist two types of attacks: noise-like signal processing and geometric distortions. Security, on the other hands, denotes the ability of a watermarking scheme to prevent hidden Contents lists available atSciVerse ScienceDirect

journal homepage:www.elsevier.com/locate/sigpro

Signal Processing

n

Corresponding author. Tel.: þ886 6 275 7575x62522; fax: þ 886 6 208 8075.

E-mail addresses: [email protected] (J.-S. Tsai), [email protected] (W.-B. Huang),

[email protected] (Y.-H. Kuo), [email protected] (M.-F. Horng).

(2)

watermarks from being accessed by unauthorized users. For the attacks to security, it is usually assumed that the unauthorized users know all knowledge about the water-marking algorithm except the secret key and they try to estimate the hidden watermarks through observing the watermarked images. The security of a watermarking scheme can be measured by analyzing the leakage of information about the hidden watermarks from observa-tions. Until now, most existing watermarking methods have been driven by improvement of robustness, includ-ing the spread spectrum [1,2], quantization index modulation [8,9], and resynchronization [10–27] schemes. But there has been little attention to security in water-marking research. However, recent studies[3–6] have shown that security is as important as robustness in developing digital watermarking schemes. Because a successful attack to security can completely break a watermarking system even though it is robust, designers should consider not only the robustness but also the degree of security in their water-marking schemes.

In this paper we use both robustness and security perspectives to investigate current feature-based water-marking methods, which are resynchronization schemes that exploit the invariant features of a medium to resist attacks. Bas et al. used the Harris detector to extract feature points from an image and the Delaunay tessella-tion to form triangle meshes with theses points for watermarking[16]. Tang and Hang adopted the Mexican Hat wavelet to extract feature regions, and exploited image normalization and FFT to hide watermarks [17]. These two methods exhibit good robustness against most attacks, but features are probably not extracted correctly after suffering from scaling attacks. Therefore, Seo et al. proposed a watermarking method based on the scale-space theory to mitigate this problem[18,21]. Robustness can be also achieved by applying the scale-invariant feature transformation[19], the Harris–Laplacian detector

[22,24], or the difference of Gaussian [23]. Recently, Gao et al.[25]used the afﬁne covariant regions to provide good resistance to geometric distortions. However, most of the existing methods usually suffer from inability to resist random or region-of-interest (ROI) cropping attacks and ineffectiveness in security-related applications.

The invalidation for cropping attacks will become more serious due to the phenomenon described in this paragraph. Since the magnitude of pixels in a feature region will be modiﬁed when a watermark is inserted into this region, it is preferred to select non-overlapping regions for watermarking to avoid major degradation of image quality. In order to obtain the non-overlapping regions, some reference parameters have been exploited in existing methods. For example, the corner response

[18,21] and the number of neighboring feature points within a region [17,22] are used to remove overlapping feature regions. In [24], the minimum spanning tree (MST) clustering algorithm was used to cluster the feature regions into groups according to a distance constraint. The region with the largest corner response in each group is then selected to be watermarked. However, non-over-lapping feature regions selected for watermarking by these parameters cannot guarantee that watermark

regions are well distributed over an image. Thus, the probability of successful cropping attacks is raised because the selected regions do not always have the maximum cover range.

Next we discuss the issue of ineffectiveness in secur-ity-related applications. It is here assumed that the attacker knows the details of watermarking algorithm except the secret key according to the Kerckhoffs’ principle. The secret key is an input to some mapping functions that outputs secret parameters, such as the watermark sequence[5]. Without knowledge of the key, the secret parameters cannot be forged or estimated. Unfortunately, it is not difﬁcult for attackers to extract the hidden watermark sequence embedded by most of the existing feature-based methods because of information leakage. The information leakage denotes the information about the hidden watermark sequence achieved from the attacker’s observation [6]. The leakage is mainly from which the watermarked feature regions’ locations and sizes are publicly known and each feature region is embedded with the same watermark. This weakness enables the security attacks to easily break down a watermarking system. For example, the attacker can apply a collusion attack to remove the hidden watermark or a copy attack to fake a watermarked image by collect-ing and analyzcollect-ing a set of feature regions with the same watermark from a watermarked image[28,29,39].

This paper proposes a novel feature-based watermark-ing method that (1) optimizes the cover range of the hidden watermarks for resisting cropping attacks and (2) enhances the security to prevent unauthorized users from accessing the secret parameters. Initially, the Harris– Laplacian detector, which uses the scale-adapted auto-correlation matrix to localize points in the scale space and invokes the Laplacian-of-Gaussian operation to select the points attaining an extremum over scales, is applied to an image to detect its feature points at multiple scale levels

[31,32]. Around these feature points, the targeted circular feature regions are determined based on their character-istic scales and a secret key. The high repeatability of these feature regions offers robustness against transla-tion, rotatransla-tion, scaling, and partial illumination changes, while secrecy of the region size makes it difﬁcult for an attacker to estimate exact range of feature region. Since these feature regions are substantially overlapped and not all stable, we propose a heuristic algorithm to select an optimal non-overlapping region set for watermarking. The region set also features a maximum distribution over the target image to tolerate the cropping attacks. The selection process further incorporates randomization to avoid an attacker correctly identifying the watermarked regions. This work is formulated as a multidimensional knapsack problem and is solved by a genetic algorithm-based procedure. Finally, we perform normalization for each selected region to obtain geometric invariance, and generate a region-dependent watermark sequence, which eliminates the problem of hiding the same signal multiple times, to be additively embedded into the spatial domain of each invariant region.

Comparing the existing feature-based methods[16–27], we investigate the important issues of feature-based

(3)

watermarking methods to improve the robustness and security. First, we propose the optimal region selection process to enhance the resistance to cropping attacks. Different from our previous work [26], the method proposed in this paper does not need the time-consuming simulated attacking procedure, and its goal is to make the selected feature regions achieve the greatest distribution over an image to withstand cropping attacks, which is not considered in [26]. Moreover, the detailed implementa-tion of the genetic algorithm-based heuristics is described in this paper. The experimental results have demon-strated that our method has better coverage in the resilience to ROI cropping, random cropping and centered cropping attacks. Also, the evaluation on the StirMark benchmark conﬁrmed the goodness of our method in the resistance to other attacks, including noise-like signal processing and geometric distortions. Second, we propose the region-dependent watermark based on feature descriptor, and incorporate randomization in feature region detection and feature region selection for security enhancement. The region-dependent watermark is derived from the framework of the content-dependent watermark with DCT block-based media hashes[39]. The novelty is that we exploit the feature description as the media hash, which is created by generating the orienta-tion histograms. The feature descriporienta-tion has been proved to be distinctive and robust[33], and it is obtained during feature detection without additional DCT operations[39]. Furthermore, the incorporated randomization in deter-mining the feature regions for watermarking makes the secrecy of their locations and sizes to mitigate the security risk. A security analysis in terms of differential entropy for feature region detection and feature region selection is given to demonstrate the effectiveness of this paper.

The rest of this paper is organized as follows. Feature detection with a controlled secret leakage is presented in

Section 2and a novel feature region selection scheme to enhance both robustness and security for watermarking is also proposed. In Section 3, the details of region-depen-dent watermark embedding and detection schemes based on local features are described. The experimental results, mainly for robustness evaluation and security analysis, are given inSection 4. Concluding remarks are drawn in

Section 5.

2. Robust and secure image features for watermarking Local features representing image structures, ranging from points to regions, have been adopted in many applications, such as object recognition, image retrieval, and camera calibration[30–34]. These features, which are powerful references, have also been applied successfully in feature-based watermarking methods since they can be preserved after suffering distortion such as scaling, rota-tion, or illumination changes. In general, a feature detector performs a speciﬁc transformation on an image to extract local features for watermark embedding and detection. However, a feature region extracted by a detector is not directly applicable to digital watermarking because of the following issues. The locations and sizes of extracted

features can be publicly found by the attackers. Embedding watermarks into all regions will also cause heavy image degradation and low robustness since most of features are overlapped. Although many new feature detectors have been proposed to enhance the robustness of feature-based watermarking[16–27], most methods are still vulnerable to security attacks and cropping attacks. Therefore, a qualiﬁed feature-based watermarking scheme should examine the robustness of the adopted feature detector, avoid the information leakage of secret parameters, and determine an appropriate non-overlapping feature region set. This section presents two processes, feature region detection and feature region selection, which are impor-tant in achieving the desired goal.

2.1. Detection of robust and secure features

In this section, the Harris–Laplacian detector, which consists of scale-adapted auto-correlation matrix and the Laplacian-of-Gaussian operation, is adopted[31,32], while the secret leakage is carefully controlled in order to identify local image features. First, the scale space of an input image I is calculated by the function L at a set of scales to represent different levels of resolutions, which is formulated as

Lðx,

s

DÞ ¼Gðx,

s

DÞn_IðxÞ _ð1Þ

where x¼(x,y) denotes the image spatial coordinate,

s

Dis

the differential scale, ‘‘n_{’’ represents the convolution}

opera-tion, and the uniform Gaussian kernel G is deﬁned by Gðx,

s

DÞ ¼ 1 2

ps

2 D eðx2_þy2_Þ=2_s D_: _ð2Þ

Then the scale-adapted auto-correlation matrix

m

(x,

s

I,

s

D) is applied in the scale space to describe the

local image structure, and it is formulated by

m

ðx,

s

I,

s

DÞ ¼

s

2DGðx,

s

IÞn L2 xðx,

s

DÞ LxLyðx,

s

DÞ LxLyðx,

s

DÞ L2yðx,

s

DÞ 2 4 3 5 ð3Þ where

s

Iis the integral scale, and Liis the ﬁrst derivative

calculated in the i direction that iA{x,y}. The corner response estimating principal curvature of the matrix is computed by its trace and determinant as

Cðx,

s

I,

s

DÞ ¼detð

m

ðx,

s

I,

s

DÞÞ0:04traceð

m

ðx,

s

I,

s

DÞÞ: ð4Þ

The feature point with large corner response repre-senting signiﬁcant curvatures has higher repeatability. The candidate points are then determined if their corner response is a local maximum and larger than a threshold TR used for ﬁltering out unstable feature regions.

However, it is difﬁcult to be set as a ﬁxed value for different input images[34]. According to the suggestion in

[43], the threshold should be set to 1% of the maximum response value of all extracted feature regions. In order to achieve scaling invariance, the integral scale of all candi-date points is compared to the characteristic scale of local image structure. The characteristic scale, which is rela-tively independent of scale change, is obtained by search-ing for a local extremum over multiple scale levels of Laplacian-of-Gaussian. Candidate points for a set of scale levels

s

n are identiﬁed by setting

s

I¼

s

n and

(4)

s

D¼0.7

s

Iwhere

s

n¼{

d

i

s

09

s

0¼1.5,

d

¼1.1, i¼1, 2, y, n}.

The scale step factor

d

between two successive levels affect the accuracy of the scale of the candidate point. It should be small to achieve high accuracy and is set to 1.1 in this paper according to the suggestion from[32]. The number of scale levels n depends on the possible scale changes of an image for different applications and is set as 15 in our experiments. The Laplacian-of-Gaussian of the candidate points is calculated as

9LoGðx,

s

nÞ9 ¼

s

2n9Lxxðx,

s

nÞ þLyyðx,

s

nÞ9: ð5Þ

A candidate point at the ith scale level is regarded as a feature point with a characteristic scale

s

c(

s

c¼

d

i

s

0) if its

Laplacian-of-Gaussian is a local extremum over all scale levels and is higher than a pre-deﬁned threshold as follows:

9LoGðx,

s

iÞ94 9LoGðx,

s

jÞ9, j 2 fi1,i þ 1g ð6Þ

9LoGðx,

s

iÞ94TLoG: ð7Þ

The threshold TLoG is set to 10, which refers to the

suggestion from[31]. In order to achieve rotation invariance and incorporate controlled secret leakage in the outputs of the feature detection, each feature point is further used as the center to derive a corresponding circular feature region with a key-dependent radius r determined

r ¼

a

U

s

_c ð8Þ

which is formulated by a secret key

a

and the feature point’s characteristic scale

s

c. Obviously, the circular feature

regions obtained by the scale-adapted auto-correlation matrix and the Laplacian-of-Gaussian operation are highly distinctive and matched with a high repeatability against various image distortions [31,32,34]. The key-dependent radius prevents an attacker from easily accessing a feature region by controlling the uncertainty of its size, and infor-mation leakage is also reduced while the watermark is embedded into the region.

2.2. Feature region selection

This work aims to obtain appropriate non-overlapping regions for watermarking since there are serious overlaps and instability in the extracted feature regions. In addi-tion to removing some overlapping regions, the feature regions selected should have maximum distribution over the image to withstand cropping attacks and to incorpo-rate randomization for security. Therefore, this work is formulated as an optimum problem constricted by image quality and regions’ overlapping circumstance as follows: maximize X NR j ¼ 1

b

jrjsj ð9Þ subject to X NR j ¼ 1 qjsjrTq ð10Þ and XNR j ¼ 1 pijsisjo1, i ¼ 1,2, . . ., NR ð11Þ

where NRis the number of feature regions extracted, rjis

the radius of region j, {

b

j} are key-dependent

pseudo-random numbers with the mean

m

and variance

s

2

, and sj

is deﬁned as sj¼

1, if the region j is selected; 0, otherwise:

(

ð12Þ The variable qj denotes the distortion of a

water-marked region j compared with its original region, and Tq refers to the limitation of quality degradation of an

image after being watermarked. Eq. (11) means that only one region can be selected in each overlapping case. The value of pijis dependent on the overlapping situation of

the two regions of i and j: pij¼

1, if the region i overlaps with region j, and iaj; 0, otherwise:

(

ð13Þ In order to solve this combinatorial optimization problem, we transform it to a multidimensional knapsack problem (MDKP) by modifying the expression of its constraints as follows: maximize X NR t ¼ 1

b

trtst ð14Þ subject to X NR t ¼ 1

o

ktstrTok, k ¼ 1,2, . . ., m: ð15Þ In Eq. (15), the variables

o

kt and Tok represent the composite weights and constraints of quality distortion and overlapping status speciﬁed in Eqs. (10) and (11), respectively. When k¼1, then

o

1t¼qtand To1¼Tq, which is equivalent to the constraint formulated by Eq. (10). Each k greater than 1 correspondingly denotes a speciﬁc index pair (i, j), iaj, in Eq. (11). Then the constraints speciﬁed in Eq. (11) can be reformulated into Eq. (15) as follows:

Referring to Eq. (11), it is obvious that

pijsisjo1, 8ri,8rj: ð16Þ

Since si, sjand pijA{0,1}, Eq. (16) can be rewritten as Eq. (17)

pijsiþpijsjr1 ð17Þ

that is,

0 s1þ0 s2þ þpijsiþ þpijsjþ þ0 sNRr1: ð18Þ Let Tok¼1, and

o

kt¼0 when tai or j, otherwise

o

kt¼pij. We can get

XNR

t ¼ 1

o

ktstrTok: ð19Þ

Aggregating Eq. (19) for different values of k, we conclude Eq. (15) and m ¼ ðN2RNRÞ=2þ 1.

Since MDKP is an NP-hard problem [36], the genetic algorithm (GA), a heuristic search approach based on the principles of evolution in nature [36,37], is employed to efﬁciently obtain a near-optimal solution in this paper. GA exploits the string structures to make an effective search, and works with a population of individuals that represent candidate solutions to a given optimization problem. It is

(5)

very likely that the obtained solution is a global solution since there are crossover and mutation operators and diverse individuals in the population being processed. The evaluation and analysis in[38]also demonstrate the solu-tion of MDKP determined by GA is the best approximasolu-tion to the global optimum among various optimization meth-ods. Firstly, a fixed length and fixed order binary bit string S 2 f0,1gNR_{representing a candidate region set, in which 1} at the jth bit indicates region j is selected, is regarded as a chromosome of an individual in a population for GA opera-tion. The length of S depends on the number of the extracted feature regions. The GA-based search procedure to find a near-optimal solution includes the following steps: 1) Population initialization: the initial population is drawn

randomly to maintain diversiﬁcation of chromosomes. Each individual in the population should be a feasible solution without violating the constraints in Eq. (15). 2) Fitness evaluation: Fitness is used to evaluate the

possibility of an individual to be the best solution. Each individual in current population has its own ﬁtness to represent degree of success as shown by FitnessðSÞ ¼X

NR

j ¼ 1

b

jrjS½j ð20Þ

where S[j] denotes the jth bit in the chromosome of an individual. The ﬁtness corresponds to the objective function in Eq. (14), and maximization of the ﬁtness leads to the best solution of feature region selection. 3) Parent selection: This step is to select individuals from a

population for a mating pool to generate new offspring. Based on the natural principle of survival of the ﬁttest, the binary tournament selection is used, which works by forming two tournament pools of individuals, each containing two individuals picked randomly from the population. Two individuals with the highest ﬁtness, each drawn from one of the two tournament pools, are selected to be parents.

4) Crossover and mutation: The generation following the selected parent individuals is obtained by two GA operators, crossover, and mutation. First, uniform crossover is adopted for any two parents to generate a single child whose chromosome is determined by copying the corresponding bits in the chromosomes of the two parents. Each copied bit is chosen randomly with equal probability from the two parents using a binary random number generator. If the random num-ber is 1, the bit is copied from the ﬁrst parent; otherwise it is copied from the second parent. Then the mutation operation is used to ﬂip a small number of bits in the child’s chromosome, changing them from 0 to 1 or vice versa. It is noted that the generated child solution by crossover and mutation operators may not be feasible due to the MDKP constraints. A repair operator is used here to overcome this problem[37]. 5) Termination: An iterative process from step 2 to this

step is executed to find the best solution. The termina-tion conditermina-tion is satisfied when either a user-defined maximum number of iterations is reached or the fittest one is unchanged during a large number of iterations.

The computational cost of the proposed feature region selection method is dominated by the GA-based search procedure for solving MDKP. Furthermore, the computa-tional cost of this procedure mainly depends on the number of feature regions and the constraints (image quality and regions’ overlapping circumstances) in MDKP. The number of feature regions is proportional to the length of individual chromosome that affects the execu-tion time in ﬁtness evaluaexecu-tion, crossover, and mutaexecu-tion. As for the constraints, they are related to the execution time for searching feasible solutions in GA. A more detailed complexity analysis of GA for solving MDKP can be found in [37]. Basically, the computational cost of feature region selection in most of the existing feature-based watermarking methods only depends on the regions’ overlapping circumstances, for example, the operations for removing region overlapping by comparing the region’s corner response[18,21] and the number of neighboring feature points within a region [17,22]. In

[24], the operational cost is to cluster the feature regions into groups by the minimum spanning tree clustering algorithm. In order to achieve the desired optimization goal, the execution time of the proposed feature region selection method is longer than the above-mentioned methods. According to our empirical study that the GA-based search procedure was coded in Borland Cþþ and executed on an Intel Core2Duo 2.4 GHz PC, the execution time spent in searching the near-optimal solu-tion is within 2 min for all test images in our experiments. 3. Proposed watermark embedding and detection schemes

The detailed procedures of watermark embedding and detection are described here, with their block diagrams depicted inFigs. 1 and 2, respectively.

3.1. Watermark embedding scheme

As shown inFig. 1, the adopted feature regions in a cover image are extracted and selected by the detector and selector described inSection 2. In the step of feature

(6)

On the other hand, a too small D makes the clustering ineffective.

4.3. Discussions

In this section, we discuss how to balance the three conﬂicting factors: robustness, capacity, and impercept-ibility in the three main processes of our method: feature detection, feature region selection, and watermark inser-tion. In the feature detection, there is a trade-off among the three factors while determining the secret key for the key-dependent radius of the feature region. A large value of the secret key would increase the capacity of the region to be watermarked, but the robustness and impercept-ibility would be decreased[22,24]. Also, a too large value would cause that the ranges of most feature regions exceed the image size. On the other hand, the value should not be too small since the capacity of feature region cannot be smaller than the watermark length. Therefore, we consider that the secret key should be large enough to make all feature regions’ sizes equal or larger than the watermark length, and should be small enough to make the ranges of most feature regions not exceed the cover range of an image in our empirical study.

In the feature region selection process, the capacity is regarded as a constant since the same watermark is embedded into each selected feature region in the feature-based watermarking methods, and its size is determined in the feature detection process. So, there is a trade-off between robustness and imperceptibility in the feature region selection. Selecting more feature regions to watermarking will produce more redundant watermarks for an image. This will increase the robust-ness but decrease the imperceptibility of the image. In the feature region selection process, we use the threshold Tq

to limit the quality degradation of the image to be watermarked and select the regions to achieve the best robustness under the limitation. It decides the level of the imperceptibility in the proposed method, and the robust-ness of a watermarked image is also determined techni-cally. For example, Tqis set as 40 dB by considering the

PSNR between an image and its watermarked image. Then, embedding watermark into the selected regions will not degrade the image quality below 40 dB.

Finally, in the watermark insertion process, there is also a trade-off, related to the watermark embedding strength for each region, between robustness and imper-ceptibility. The capacity is still considered as a constant since the region’s size is determined in the feature detection process. A large embedding strength will increase the robustness but decrease the imperceptibility. Here, we use the noise visibility function (NVF) to deter-mine the appropriate embedding strength, which can avoid large degradation of image quality. The NVF that characterizes the local image properties can achieve the best balance for the two factors[42].

5. Conclusions

In this paper, we develop a novel method to jointly enhance the robustness and security of feature-based

image watermarking schemes. The controlled randomiza-tion is incorporated in determining the feature regions of an image for mitigating the leakage of secret information. In addition, an optimal selection process is proposed, formulated as the multidimensional knapsack problem and solved by genetic algorithm-based heuristics. The experimental results of robustness evaluation demon-strate that our method can effectively resist various attacks, including noise-like signal processing and geo-metric distortions. The security evaluation in terms of differential entropy is also derived, and the performance of the proposed method is conﬁrmed.

Acknowledgments

The authors would like to thank anonymous reviewers for giving constructive and useful comments to improve this paper. This work is supported in part by the National Science Council of Taiwan under Grants 97-2221-E-006-144-MY3 and 98-2221-E-006-222-MY3.

References

[1] I.J. Cox, J. Kilian, F.T. Leighton, T. Shamoon, Secure spread spectrum watermarking for multimedia, IEEE Transactions on Image Proces-sing 6 (12) (1997) 1673–1687.

[2] I.J. Cox, M.L. Miller, J.A. Bloom, Digital Watermarking, Morgan Kaufmann, San Francisco, CA, 2001.

[3] F. Cayre, C. Fontaine, T. Furon, Watermarking security: theory and practice, IEEE Transaction on Signal Processing 53 (10) (2005) 3976–3987.

[4] F. Cayre, P. Bas, Kerckhoffs-based embedding security classes for WOA data hiding, IEEE Transaction on Information Forensics Security 3 (1) (2008) 1–15.

[5] L. Pe´rez-Freire, P. Comesan˜ a, J.R. Troncoso-Pastoriza, F. Pe´rez-Gon-za´lez, Watermarking security: a survey, Transactions on Data Hiding and Multimedia Security I 4300 (2006) 41–72.

[6] L. Pe´rez-Freire, F. Pe´rez-Gonza´lez, Spread-spectrum watermarking security, IEEE Transaction on Information Forensics Security 4 (4) (2009) 2–24.

[7] X. Gao, L. An, Y. Yuan, D. Tao, X. Li, Lossless data embedding using generalized statistical quantity histogram, IEEE Transactions on Circuits and Systems for Video Technology 21 (8) (2011) 1061–1070.

[8] B. Chen, G. Wornell, Quantization index modulation: a class of provably good methods for digital watermarking and information embedding, IEEE Transactions on Information Theory 47 (4) (2001) 1423–1443.

[9] J. Eggers, R. Ba ¨uml, R. Tzschoppe, B. Girod, Scalar costa scheme for information embedding, IEEE Transaction on Signal Processing 51 (4) (2003) 1003–1019.

[10] J. O´ Ruanaidh, T. Pun, Rotation, scale and translation invariant spread spectrum digital image watermarking, Signal Processing 66 (3) (1998) 303–317.

[11] C.Y. Lin, M. Wu, J.A. Bloom, I.J. Cox, M.L. Miller, Y.M. Lui, Rotation, scale and translation resilient watermarking for image, IEEE Trans-actions on Image Processing 10 (5) (2001) 767–782.

[12] P. Dong, J.G. Brankov, N.P. Galatsanos, Y. Yang, F. Davoine, Digital watermarking robust to geometric distortions, IEEE Transactions on Image Processing 14 (12) (2005) 2140–2150.

[13] D. Simitopoulos, D.E. Koutsonanos, M.G. Strintzis, Robust image watermarking based on generalized radon transformations, IEEE Transactions on Circuits and Systems for Video Technology 13 (8) (2003) 732–745.

[14] M. Kutter, Watermarking resisting to translation, rotation and scaling, Proceedings of the SPIE Multimedia Systems and Applica-tions 3528 (1998) 423–431.

[15] S. Pereira, T. Pun, Robust template matching for afﬁne resistant image watermarks, IEEE Transactions on Image Processing 9 (6) (2000) 1123–1129.

(7)

[16] P. Bas, J.M. Chassery, B. Macq, Geometrically invariant watermark-ing uswatermark-ing feature points, IEEE Transactions on Image Processwatermark-ing 11 (9) (2002) 1014–1028.

[17] C.W. Tang, H.M. Hang, A feature-based robust digital image water-marking scheme, IEEE Transaction on Signal Processing 51 (4) (2003) 950–959.

[18] J.S. Seo, C.D. Yoo, Localized image watermarking based on feature points of scale-space representation, Pattern Recognition 37 (7) (2004) 1365–1375.

[19] H.Y. Lee, H. Kim, H.K. Lee, Robust image watermarking using local invariant features, Journal of SPIE Optical Engineering 45 (3) (2006) 037002-1–037002-11.

[20] C.S. Lu, S.W. Sun, C.Y. Hsu, P.C. Chang, Media hash-dependent image watermarking resilient against both geometric attacks and estimation attacks based on false positive-oriented detection, IEEE Transactions on Multimedia 8 (4) (2006) 668–685.

[21] J.S. Seo, C.D. Yoo, Image watermarking based on invariant regions of scale-space representation, IEEE Transaction on Signal Processing 54 (4) (2006) 1537–1549.

[22] X. Wang, J. Wu, P. Niu, A new digital image watermarking algorithm resilient to desynchronization attacks, IEEE Transaction on Information Forensics Security 2 (4) (2007) 633–655. [23] D. Zheng, S. Wang, J. Zhao, RST invariant image watermarking

algorithm with mathematical modeling and analysis of the water-marking processes, IEEE Transactions on Image Processing 18 (5) (2009) 1055–1068.

[24] C. Deng, X. Gao, X. Li, D. Tao, Local histogram based geometric invariant image watermarking, Signal Processing 90 (12) (2010) 3256–3264.

[25] X. Gao, C. Deng, X. Li, D. Tao, Geometric distortion insensitive image watermarking in afﬁne covariant regions, IEEE Transactions on Systems Man and Cybernetics C Applied Review 40 (3) (2010) 278–286.

[26] J.S. Tsai, W.B. Huang, Y.H. Kuo, On the selection of optimal feature region set for robust digital image watermarking, IEEE Transactions on Image Processing 20 (3) (2011) 735–743.

[27] J.S. Tsai, W.B. Huang, C.L. Chen, Y.H. Kuo, A feature-based digital image watermarking for copyright protection and content authen-tication, in: Proceedings of the IEEE International Conference on Image Processing, 2007, pp. 469–472.

[28] S. Voloshynovskiy, S. Pereira, V. Iquise, T. Pun, Attack modeling: towards a second generation watermarking benchmark, Signal Processing 81 (6) (2001) 1177–1214.

[29] F. Deguillaume, S. Voloshynovskiy, T. Pun, Secure hybrid robust watermarking resistant against tampering and copy attack, Signal Processing 83 (10) (2003) 2133–2170.

[30] C. Harris, M. Stephen, A combined corner and edge detector, in: Proceedings of the Fourth Alvey Vision Conference, 1988, pp. 147–151. [31] K. Mikolajczyk, C. Schmid, Indexing based on scale invariant interest points, in: Proceedings of the Eighth International Con-ference on Computer Vision, 2001, pp. 525–531.

[32] K. Mikolajczyk, C. Schmid, Scale and afﬁne invariant interest point detectors, International Journal of Computer Vision 60 (1) (2004) 63–86.

[33] D.G. Lowe, Distinctive image features from scale-invariant keypoints, International Journal of Computer Vision 60 (2) (2004) 91–110. [34] K. Mikolajczyk, T. Tuytelaars, C. Schmid, A. Zisserman, J. Matas,

F. Schaffalitzky, T. Kadir, L.V. Gool, A comparison of afﬁne region detectors, International Journal of Computer Vision 65 (1–2) (2005) 43–72.

[35] A.P. Fabien, Petitcolas, Watermarking schemes evaluation, IEEE Signal Processing Magazine 17 (5) (2000) 58–64.

[36] H. Kellerer, U. Pferschy, D. Pisinger, Knapsack Problems, Springer, Berlin, 2004.

[37] P.C. Chu, J.E. Beasley, A genetic algorithm for the multidimensional knapsack problem, Journal of Heuristics 4 (1) (1998) 63–86. [38] R.J. Moraga, G.W. DePuy, G.E. Whitehouse, Meta-RaPS approach for

the 0–1 multidimensional knapsack problem, Computers and Industrial Engineering 48 (1) (2005) 83–96.

[39] C.S. Lu, C.Y. Hsu, Near-optimal watermark estimation and its countermeasure: antidisclosure watermark for multiple watermark embedding, IEEE Transactions on Circuits and Systems for Video Technology 17 (4) (2007) 454–467.

[40] A. Swaminathan, Y. Mao, M. Wu, Robust and secure image hashing, IEEE Transaction on Information Forensics Security 1 (2) (2006) 215–230.

[41] T.M. Cover, J.A. Thomas, Elements of Information Theory, 2nd edition, John Wiley & Sons, New York, 1991.

[42] S. Voloshynovskiy, A. Harrigel, N. Baumgartner, T. Pun, A stochastic approach to content adaptive digital image watermarking, in: Proceedings of the International Workshop on Information Hiding, LNCS 1768, 1999, pp. 211–236.

[43] C. Schmid, R. Mohr, C. Bauckhage, Evaluation of interest point detectors, International Journal of Computer Vision 37 (2) (2000) 151–172.

[44] S. Ehsan, N. Kanwal, A.F. Clark, K.D. McDonald-Maier, Measuring the coverage of interest point detectors, in: Proceedings of the International Conference on Image Analysis Recognition, LNCS 6753, 2011, pp. 253–261.

[45] G. Schaefer, M. Stich, UCID—an uncompressed colour image data-base, in: Proceedings of the SPIE Storage and Retrieval Methods and Applications for Multimedia, 2004, pp. 472–480.