Elsevier

Information Fusion

Volume 8, Issue 4, October 2007, Pages 354-365
Information Fusion

Genetic perceptual shaping: Utilizing cover image and conceivable attack information during watermark embedding

https://doi.org/10.1016/j.inffus.2005.09.007Get rights and content

Abstract

We describe a new watermarking scheme based on intelligent shaping of a digital watermark using Genetic Programming (GP). The proposed method, in addition to achieving a superior tradeoff between watermark robustness and imperceptibility, is also able to structure the watermark in accordance with an anticipated attack. This has been achieved by simultaneously hiding the watermark as well as spreading and fusing it in such a way to resist the conceivable attack. Robustness versus imperceptibility tradeoff and increase in bit correct ratio after attack, have been employed as the optimization criteria in the GP search. The concept of bonus fitness has been used to implement multi-objective fitness based GP evolution. Experiments on standard images indicate that such watermark shaping functions could be developed that are cover image independent and enhance imperceptibility. They offer high resistance against removal and interference attacks of Checkmark benchmark.

Introduction

Due to the rapid growth in the use of digital media, there is an increasing concern about unauthorized handling, copying and reuse of information. Watermarking which is being considered as the practice of imperceptibly altering data to embed a message about that data, is an effective way to counter these types of problems [1]. Digital watermarking is performed upon a variety of different digital materials, like audio, images, text, movies and 3D models. It has also a broad range of applications, like ownership assertion, authentication, broadcast monitoring, and integrity control [2]. In a watermarking system, there is an intrinsic relation between two of its most important, but contradicting properties: robustness and imperceptibility. Imperceptibility means that the watermarked data should be perceptually equivalent to the original, unwatermarked data. On the other hand, robustness means that the watermark should not be rendered undetectable, unless damaging the usefulness of the cover data itself [3]. If we try to improve the watermark imperceptibility, robustness decreases and vice versa. Consequently, one needs to make a tradeoff according to the application domain. For this purpose, different methods, both in spatial as well as transformed domain, have been used to tailor a watermark according to the cover image [4], [5], [6], [7], [8], [9], [10], [11].

Watermarks are rendered undetectable with an attack, where the attack is defined as any processing of the watermarked data that might damage the watermark [1], [3]. Thus watermarking can be viewed as a reliable mode of communication to transfer important information (i.e. a watermark) embedded in a signal (e.g. a cover image) safely through a hostile environment [12]. Attacks can be intentional such as watermark estimation using Wiener filtering or unintentional such as JPEG compression. An extensive list of attacks appears in [1], [13], [14], [15], [16], [17].

Due to the nature of diverse types of attacks, there is no generic watermarking scheme that could resist all sorts of attacks. However, it can be assumed that many applications are not concerned with all conceivable attacks, but with specific attacks that might occur before decoding [1]. Investigators have addressed this problem in various ways. One way is to develop watermarking approaches suitable for the anticipated attack [18]. For example, in case of rotational attack, alteration in the phase, rather than the amplitude of the Fourier component, is performed to embed a watermark [19]. Another possibility is to achieve robustness against the probable processing of the watermarked image, by restructuring the watermark. In this scenario, robustness is often achieved at the expense of imperceptibility, computational cost, data payload, or even robustness to some other processing.

To defend attacks, efforts have been made to increase robustness at low cost of imperceptibility. For instance Jonathan et al. [3] have taken a theoretical approach to answer the complex question of “how should a watermark be structured to maximize its robustness”. They have proposed that the watermark power spectrum should be proportional to that of the original signal. Liang et al. [21] propose robust watermarking using robust coefficients for embedding. Huang et al. [8], [20], on the other hand, have used Genetic Algorithms for the selection of coefficients to be altered for watermark embedding. However, these efforts concentrate on tailoring just the choice of specific coefficients, not the whole watermark, to a cover image and intended attack. In fact, they are not using perceptual models; rather a fixed strength of the alteration is used for each selected DCT coefficient.

Perceptual models [23], [24], [25], [26], as those of Watson’s, which have been frequently used in image compression are used to compute the strength of the alteration for each selected coefficient. These perceptual models make a tradeoff between robustness and imperceptibility according to the cover image. However, they do not take into consideration the watermark application and thus the intended attacks. For instance, when the watermarked image is expected to be JPEG compressed, it is judicious to structure the watermark in view of the JPEG compression. Pertinent examples exist in literature [27], where appropriate watermarking approaches as well embedding domains have been studied to achieve robustness against JPEG compression.

One way to restructure a watermark in view of the anticipated attack is to keep high watermark strength for those selected coefficients that are less affected by the attack. However, firstly this requirement needs to consider limitations imposed by imperceptibility. Secondly, this requirement varies for different types of attacks. Consequently, our aim in this work is to propose and study an automatic system that can restructure the watermark in accordance to the cover image and intended attack. Specifically, to propose a system for developing suitable watermark shaping functions, which are image independent and intended attack-resistant.

We address these requirements through the following contributions:

  • 1.

    We consider the perceptual shaping of a watermark to be vital, not only for imperceptibility enhancements, but we realize it to be a method of structuring the watermark in accordance to the anticipated attack.

  • 2.

    We introduce the concept of developing complex and appropriate watermark shaping functions from the existing ones. Specifically, we consider Watson’s perceptual model, characteristics of the HVS and information about the distortion caused by the anticipated attack, as independent variables and genetically search for application-specific watermark shaping functions.

The idea used is analogous to combining classifiers for developing complex, but appropriate classifier for a certain application of pattern recognition [28]. We call this technique as Genetic Watermark Shaping Scheme (GWSS) and the genetically developed watermark shaping functions as Genetic Watermark Shaping Functions (GWSF).

In Section 2, we discuss perceptual shaping of a digital watermark including discussion about perceptual models. We discuss attacks and their countermeasures in Section 3, while imperceptibility and robustness measures in Section 4. Section 5 explains our proposed technique GWSS. This includes description of various modules of GWSS and explains our bonus fitness idea used in the multi-objective based GP evolution. This section also describes the testing and comparison phase of the evolved GWSF. Section 6 presents implementation details and section 7 gives results and discussions. Conclusion and future work are discussed at the end.

Section snippets

Perceptual shaping of a digital watermark

A watermark is generally embedded in a cover image with a high strength in areas where it is well hidden and with a low strength in places where it is clearly discernible. This type of strategy is known as perceptual shaping of a watermark [1]. For this purpose, usually perceptual models that are used in compression techniques are employed. These perceptual models are able to learn the content of a cover image by exploiting the sensitivities/insensitivities of an HVS. They take advantage of

Attacks and their countermeasures

Digital watermarks can be attacked in a variety of different ways and each application requires its own type of robustness. Cox et al. [1] have discussed in detail the types and levels of robustness that might be required for a particular watermarking application. They have discussed some of the attacks as well as their countermeasures. Voloshynovsky et al. [13] have classified attacks into four basic categories: removal and interference attacks, geometrical attacks, cryptographic attacks and

Watermark robustness and imperceptibility measures

The imperceptibility of a watermark is generally measured in terms of weighted Peak Signal to Noise Ratio (wPSNR) [15], Watermark to Document Ratio (WDR) [31] and Structural Similarity Index Measure (SSIM) [32]. SSIM measure uses the hypotheses that HVS is highly adopted for extracting structural information. It is argued that natural image signals are highly structured, as the nearby pixel exhibit strong dependencies [32]. These dependencies provide information about the structure of the

Proposed technique for developing a GWSF

The basic architecture of our proposed scheme for developing GWSF is shown in Fig. 1. Five modules work in a cyclic fashion. We first explain the overall working of the basic architecture. Details of the individual modules are given in Section 5.1.

The GP module produces a population of GWSF. Each GWSF is presented to the perceptual shaping module, where it is applied to the cover image in DCT-domain, generating a perceptual mask. In the watermarking stage, the watermark is shaped using the

Implementation details

We have used MATLAB environment for our experimental studies. To employ GP, we use GPLAB toolbox [36], [37]. The GP parameter settings are shown in Table 1, while the remaining parameters are used as default in the software.

Lena image of size 256 × 256 is used as a cover image with Nd = 22 (7–29 in zig-zag order) during the GP simulation. Message size is kept equal to 64 bits. Following [23], [24], the parameters of WPM are set as r = 0.7, Tmin = 1.1548, u = 1.728, fmin = 3.68 cycles/degree and aT = 0.649.

Perceptual shaping using GWSF

In Fig. 7, watermarking strength corresponding to each bandpass DCT coefficient of block-based DCT is shown. These strengths are produced by the Wiener attack-resistant GWSF for Lena image. It is observed that depending upon the current AC and DC coefficient, it provides suitable imperceptible alterations according to the spatial content of that block. This fact indicates that GWSF is able to exploit HVS for shaping the watermark according to any cover image. In other words, GWSF makes the

Conclusions

In this paper we have considered the GP-based perceptual shaping of a digital watermark in accordance to the cover image and anticipated attack. The GP tuned GWSFs are image adaptive and the GWSS as a whole is attack adaptive. A significant improvement in resistance against the intended attack is achieved by letting the GP search exploit the attack information. This is in essence, like attack-informed embedding. Both these attributes of a GWSF; superior tradeoff and high resistance against an

Acknowledgement

The authors greatly acknowledge the financial support provided by the Higher Education Commission, Government of Pakistan, under the indigenous PhD scholarship program (No. 17-6 (183)/Sch/2001).

References (39)

  • S. Voloshynovskiy, A. Herrigel, N. Baumgaetner, T. Pun, A stochastic approach to content adaptive digital image...
  • Ki-Ryong Kwon, Ji-Hwan Park, Eung-Joo Lee, Ahmed H. Tewfik, Highly reliable stochastic perceptual watermarking model...
  • P. Meerwald, A. Uhl, A survey of wavelet-domain watermarking algorithms, in: Proceedings of SPIE, Security and...
  • D. Kundur et al.

    Towards robust logo watermarkin using multiresolution image fusion principles

    IEEE Trans. Multimedia

    (2004)
  • A. Sequeira et al.

    Communication and information theory in watermarking: a survey, multimedia systems and applications IV

  • M. Kutter et al.

    A fair benchmark for image watermarking systems, electronic Imaging ’99

    (1999)
  • S. Voloshynovskiy et al.

    Attacks on digital watermarks: classification, estimation based attacks, and benchmarks

    IEEE Commun. Mag.

    (2001)
  • F. Hartung, J.K. Su, B. Girod, Spread spectrum watermarking: malicious attacks and counterattacks, in: Proc. SPIE,...
  • E. Praun, H. Hoppe, A. Finkelstein, Robust Mesh Watermarking, in: Proceedings of SIGGRAPH 1999, Computer Graphics...
  • Cited by (0)

    View full text