Watermarking with Wavelets: Simplicity Leads to Robustness
Evelyn Brannock Computer Science Department Georgia State University Atlanta, Georgia, USA 30303
[email protected]
Michael Weeks Computer Science Department Georgia State University Atlanta, Georgia, USA 30303
[email protected]
Robert Harrison Computer Science Department Georgia State University Atlanta, Georgia, USA 30303
[email protected]
1. Introduction
Abstract
The protection of creative digital content poses a difficult challenge to those who wish to retain control of their work. Historically, U.S. copyright law has tried to provide protection for this intellectual property. The improvements in digital technologies continue to ease and increase the ways in which consumers can innocently use and enjoy creative content, for example, by copying music files from a CD to store on a computer or portable music device. However, many cases are not so innocent and the media in digital form are flawlessly and inexpensively reproduced in great volumes and instantaneously distributed worldwide for more nefarious purposes such as digital piracy for profit. Therefore, creators and owners of the work are concerned that unauthorized copying and redistribution of their copyrighted works will cause their economic returns to decline. Consequently, the study of technological approaches to solve this problem has become increasingly important and popular. One such approach is digital watermarking. Watermarks serve to identify the source of the content and thus aid in investigating abusive duplication. The impact of the size and nature of the data on the robustness of the embedded watermark will be investigated, in an extension of [1]. An uncomplicated key will be used. The size of the key will not have any effect on the visibility of the watermark, but as in other cryptographic systems, for commercial applications it should be large enough to make extensive search attacks more
With ubiquitous computing comes the access of copyrighted work across computing platforms. One may have the same image (or video) on an iPod, as well as a laptop and desktop computer. Safeguarding creative content and intellectual property in a digital form has become increasingly difficult as technologies, such as the Internet, broadband availability and mobile access, advance. It has grown to be progressively easier to copy, modify and redistribute digital media, resulting in great declines in business profits. Digital watermarking is a steganographic technique that has been proposed as a possible solution to this problem. This paper examines a technique for digital watermarking which utilizes properties of the Discrete Wavelet Transform (DWT). The digital watermarking algorithm uses a database of multiple images with various properties. Eight families of wavelets, both orthogonal and biorthogonal, are compared for efficacy. To objectively measure the success of the algorithm and the influence of the mother wavelet, the PSNR for each wavelet family and image is obtained. Noise is introduced to simulate various attacks. Objective measures are used to determine the performance of the algorithm. We find that the simpler wavelet transforms, e.g. the Haar wavelet, outperform the more complex ones.
978-1-4244-1884-8/08/$25.00 ©2008 IEEE
587
Authorized licensed use limited to: CHONGQING UNIVERSITY. Downloaded on March 21, 2009 at 02:00 from IEEE Xplore. Restrictions apply.
difficult. All of the test images are grayscale. In this paper, the next section will cover the background and digital watermarking principles, section 3 will cover wavelets, and then section 4 will discuss the method used. Section 5 presents results, and section 6 concludes the paper.
(also asymmetric marking or private), in which the watermark is embedded in the original host, and is intentionally visible to the human observer. The original data is required for watermark extraction [1] [2]. The blind watermark has many more applications than the visible watermark. The subject of this paper is a blind watermark. There are various watermarking applications for images, as listed below [2] [5] [7].
2. Digital Watermarking 2.1. Definition
• Copyright protection is probably the most common use of watermarks today. Copyright owner information is embedded in the image in order to prevent others from alleging ownership of the image.
According to Hartung and Kutter, a digital watermark is “a digital code unremovably, robustly, and imperceptibly embedded in the host data and typically contains information about origin, status, and/or destination of the data” [2]. It is a form of steganography, because it hides the embedded data, often without the knowledge of the viewer or user. Since the purpose of steganography is the secret communication between two persons, the watermark can be considered to have been successfully attacked if its existence is determined. When contrasting with steganography, watermarks add the property of robustness, which is the ability to withstand most common attacks [1]. Attacks usually include two types: removing the watermark and rendering the watermark undetectable [3]. Attack categorization may include (but are not limited to) [3] [4] [5]:
• The fingerprint embeds information about the legal receiver in the image. This involves embedding a different watermark into each distributed image and allows the owner to locate and monitor pirated images that are illegally obtained. • Prevention of unauthorized copying is accomplished by embedding information about how often an image can be legally copied. An ironic example in which the use of a watermark might have prevented the wholesale pilfering of an image is in the ubiquitous “Lena” image, which has been used without the original owner’s permission.
• Adding noise such as Gaussian.
• In an image authentication application the intent is to detect modifications to the data. The characteristics of the image, such as its edges, are embedded and compared with the current images for differences.
• Using linear filtering such as low-pass filtering. • Compressing the image, such as JPEG does. • Applying transforms such as translation, rotation and scaling.
2.3. Requirements
• Permuting the original signal by rerecording or recapturing so that extracting the watermark is nearly impossible.
Obviously, an implicit requirement for a blind watermark is that it is invisible to the naked eye and should look indistinguishable from the original. There are also other requirements for successful watermarking techniques. Literature lists the following common requirements: robustness, imperceptible to statistical methods, recovery with or without the original data, extraction or verification of a given watermark, security issues and use of keys, speed, and capacity [1] [2] [4] [7]. How can each of these requirements be scored and evaluated? The ideal would be to gather a large sample of people to view the original host image, the watermark, the host image containing the embedded watermark, and the extracted watermark under excellent circumstances (good lighting, no distractions, etc.) because evaluation of the watermark involves the subjective judgment of the distortion introduced through the
Even though the rise in popularity as a research topic does not appear to have begun until the early 1990’s, the watermark has a long and distinguished history. The oldest watermarked paper has been dated back to the 13th century when papermaking artisans needed to protect their provenance [6]. The analogy can be seen to today’s watermarks; only the media is different.
2.2. Types and Applications There are two main types of watermarks. A blind (or public) watermark is invisible, and is extracted “blindly” without knowledge of the original host image or the watermark itself. The second is non-blind
588
Authorized licensed use limited to: CHONGQING UNIVERSITY. Downloaded on March 21, 2009 at 02:00 from IEEE Xplore. Restrictions apply.
process. (In general, there is a trade-off between watermark robustness, watermark perceptibility and watermark payload) [8]. However, since this was not possible for the paper, the Peak Signal-to-Noise Ratio was utilized. The quality of an N × M host image ( f (x, y)) is compared to the image containing the watermark (g(x, y)) using the formula,
regions allow us to increase the robustness of our watermark, at little to no additional impact on image quality [10]. The fact that the DWT is a multi-scale analysis can be used to the watermarking algorithm’s benefit. Multi-resolution is the process of taking a filter’s output and putting through another pair of analysis filters. The first approximation will be used as a “seed” image and recursively apply the DWT a second and third time (or however many times it is necessary to perform to find all of the areas of interest) [9]. See [14] for more background on wavelets, and [13] for wavelet history.
max pixel value PSNR(g, f ) = 20 × log10 2 ∑x,y ( f (x,y)−g(x,y)) size
2.4. Watermarking Techniques All watermarking methods share a watermark embedding system and a watermark extraction system [1] [2]. There are two main watermarking techniques available: spatial domain and frequency domain. The technique used in this paper will embed the watermark using the Discrete Wavelet Transform, utilizing a frequency domain method.
4. Implementation A simple watermark raster bitmap image was embedded in each of the images. It was created by cutting and pasting the word “Watermark” from a Microsoft Word document into Adobe Photoshop. This exhibits the use of text, such as the owner’s name, as a watermark image.
3. Wavelets The Discrete Wavelet Transform (DWT) is currently used in a wide variety of signal processing applications, such as in audio and video compression, removal of noise in audio, and the simulation of wireless antenna distribution. Wavelets have their energy concentrated in time and are well suited for the analysis of transient, time-varying signals. Since most of the reallife signals encountered are time varying in nature, the Wavelet Transform suits many applications very well [9]. We use the DWT to implement a simple watermarking scheme. The 2-D discrete wavelet transform (DWT) decomposes the image into sub-images, 3 details and 1 approximation. The approximation looks just like the original; only on 1/4 the scale. The 2-D DWT is an application of the 1-D DWT in both the horizontal and the vertical directions. The DWT separates an image into a lower resolution approximation image (LL) as well as horizontal (HL), vertical (LH) and diagonal (HH) detail components. The low-pass and high-pass filters of the wavelet transform naturally break a signal into similar (lowpass) and discontinuous/rapidly-changing (high-pass) sub-signals. The slow changing aspects of a signal are preserved in the channel with the low-pass filter and the quickly changing parts are kept in the high-pass filter’s channel. Therefore we can embed high energy watermarks in the regions that human vision is less sensitive to, such as the high resolution detail bands (LH, HL, and HH). Embedding watermarks in these
Figure 1. Watermark bitmap embedded There are numerous wavelets to choose from. Out of the box, MATLAB offers a plethora of choices [11]. This paper focuses on the question: Is there a best selection that can be made for an application? Since a unique answer to this question was not discovered, eight differing popular wavelets were used to implement for comparison and contrast; the Haar wavelet, the orthogonal, 4-coefficient, Daubechies wavelet (e.g. db2), the 32nd order Daubechies wavelet (e.g. db32), three biorthogonal wavelets, including a reverse biorthogonal wavelet (bior2.2, bior5.5, rbio6.8), the symlet 8coefficient wavelet and the 4th order coiflet wavelet.
4.1. Images Used Varying sizes, complexity and types of images were chosen. All images are grayscale. The image database consists of: • “Barbara” - 512 x 512 pixels • “Boat” - 512 x 512 pixels [15] • “Box” - 256 x 256 pixels • “Cameraman” - 256 x 256 pixels (available at [15]) • “Cell” - 190 x 158 pixels
589
Authorized licensed use limited to: CHONGQING UNIVERSITY. Downloaded on March 21, 2009 at 02:00 from IEEE Xplore. Restrictions apply.
• “Circuit” - 272 x 280 pixels
• “Filopodia” - 640 x 480 pixels (thanks to Dr. Vincent Rehder at Georgia State University)
horizontal and vertical details is found, and if that correlation exceeds a threshold (the mean of the correlation), a pixel in the watermark is located. yi is a candidate pixel of the watermark and M is the length of the watermark. 1 z= ∑ Wi∗ × yi [12] M 1,M
• “France” - 672 x 496 pixels (a converted PowerPoint slide [15])
Finally the extracted watermark is written and the PSNR is calculated.
• “Dog on Porch” - 256 x 256 pixels [9] • “Eight Coins” - 242 x 308 pixels
• “Frog” - 620 x 498 pixels [15]
5. Results
• “Gold Hill” - 512 x 512 pixels [15] The following images (figure 2 and 3) are the results of the watermarking implementation in this project. For brevity, the watermarked images have been resized from their original size and not all of the recovered watermarks are pictured. Twenty pristine original images, with eight wavelet families, and three types of noise introduced results in 640 watermarked images and 640 recovered watermarks. The watermarks in figure 2 are all from the Lena image, encoded from left to right with the Haar wavelet, Daubechies-4 coefficient wavelet, and the biorthogonal 2.2 wavelet. In figure 3, we see the recovered watermarks from the Frog image [15], with Gaussian noise, and encoded from left to right with the Haar wavelet, the biorthogonal 5.5, and the symlet 8 coefficient wavelet. Clearly, the watermarks encoded with the Haar wavelet were the best ones recovered. These recovered watermarks are representative; each image in the database had corresponding visual results.
• “Lena” - 512 x 512 pixels • “Mandrill” - 512 x 512 pixels (available at [15]) • “Moon” - 358 x 536 pixels • “Mountain” - 640 x 580 pixels [15] • “MRI” - 128 x 128 pixels • “Peppers” - 512 x 512 pixels [15] • “Pout” - 240 x 290 pixels • “Zelda” - 512 x 512 pixels [15]
4.2. Algorithm The 2-D DWT is applied to the image, giving four quadrants 1/4 the size of the original image, and producing the two matrices of coefficients that will be manipulated, the horizontal details (HL) and vertical details (LH). A pseudo random noise pattern is generated using the secret key image as a seed, and each of the bits of the watermark are embedded in the horizontal (HL) and vertical (LH) coefficient sub-bands using this pattern. The equation used to embed the watermark is: Wi = Wi + α Wi xi Wi = Wi
Figure 2. Watermark Images Recovered from Lena: Haar, DB2, and Bior2.2
for all pixels in LH, HL
for all pixels in HH, LL[12]
Wi is the watermarked image, Wi is the original image, and α is a scaling factor. Increasing α increases the robustness of the watermark, but decreases the quality of the watermarked image. We use the same α (the constant 2 on a scale of 1 to 5) as used in [12]. Finally, we write the image, and calculate the PSNR. To extract the watermark, we apply the 2-D inverse DWT to the possibly corrupted watermarked image Wi∗ . The same secret key is used to seed the random function and to generate the pseudo random noise pattern. The correlation, z, between the pseudo random noise and the
Figure 3. Watermark Images Recovered from Frog (with Gaussian noise): Haar, Bior5.5, and Symlets8 The five tables shown represent the statistics obtained. The first table examines the watermark embedding process, showing the average, minimum and maximum PSNRs. The second table shows, for the entire image database, the average PSNR obtained for each type of noise, including Gaussian, salt and pepper, and
590
Authorized licensed use limited to: CHONGQING UNIVERSITY. Downloaded on March 21, 2009 at 02:00 from IEEE Xplore. Restrictions apply.
Table 1. Watermark Embedding Wavelet Haar Daubechies Daubechies 32 Bior 2.2 Bior 5.5 Symlets 8 Coiflets 4 Rev. Bior 6.8 All
Average PSNR 25.5674 23.2076 20.0191 23.5340 24.3510 23.7301 23.8802 23.7707 23.5075
Minimum PSNR 25.1307 17.1497 20.8266 18.0592 19.0025 18.2345 18.5406 18.9390 17.1497
Table 3. Sample Data for Watermark Extraction
Maximum PSNR 28.1656 26.3296 26.9422 27.8657 26.4089 26.5825 26.7399 26.5306 28.1656
Wavelet Haar Daubechies Daubechies 32 Bior 2.2 Bior 5.5 Symlets 8 Coiflets 4 Rev. Bior 6.8 Haar Daubechies Daubechies 32 Bior 2.2 Bior 5.5 Symlets 8 Coiflets 4 Rev. Bior 6.8 Haar Daubechies Daubechies 32 Bior 2.2 Bior 5.5 Symlets 8 Coiflets 4 Rev. Bior 6.8 Haar Daubechies Daubechies 32 Bior 2.2 Bior 5.5 Symlets 8 Coiflets 4 Rev. Bior 6.8
Table 2. Watermark Embedding with Noise Salt and Speckle Gaussian Pepper Average Average Average Wavelet PSNR PSNR PSNR Haar 25.56 25.57 25.65 Daubechies 18.15 18.02 16.73 Daubechies 32 19.56 19.40 18.22 Bior 2.2 20.06 19.93 18.71 Bior 5.5 18.83 18.72 17.54 Symlets 8 19.08 18.96 17.78 Coiflets 4 19.19 19.08 17.89 Rev. Bior 6.8 19.20 19.08 17.91
speckled. The third displays data for a small sample of the images for the watermark recovery process. The fourth table examines the watermark extraction results, without noise, and the fifth table adds noise into the amalgamation. PSNR values at least 25 dB are, theoretically, the least perceptible to the human eye.
Image Barbara
Box
Cameraman
Cell
PSNR 6.11 5.82 5.94 6.02 5.92 5.93 5.93 5.94 6.10 5.87 5.94 5.98 5.93 5.93 5.93 5.94 6.06 5.86 5.94 5.98 5.93 5.94 5.94 5.94 6.01 5.88 5.94 5.97 5.93 5.94 5.94 5.94
Image Frog
Lena
Mandrill
Moon
PSNR 6.11 5.81 5.94 6.04 5.92 5.93 5.94 5.95 6.11 5.80 5.94 6.03 5.92 5.93 5.94 5.95 6.11 5.82 5.94 6.02 5.92 5.93 5.94 5.94 6.11 5.82 5.93 6.04 5.91 5.93 5.93 5.94
Table 4. Watermark Extraction
First, the watermark is embedded in the original image. No noise is introduced into the image. To simulate effects of such innocent problems such as transmission errors, or perhaps alterations of the image for other more nefarious reasons, three types of noise are then independently applied to the image to simulate image corruption. The Gaussian white noise added had a zero mean noise with 0.01 variance. The salt and pepper noise had noise density of 0.04, affecting approximately 4% of the pixels. Lastly, the speckle added multiplicative noise that is uniformly distributed random noise with mean 0 and variance 0.04. Then the watermark is extracted from the noisy image. The results are shown in the tables.
Wavelet Haar Daubechies Daubechies 32 Bior 2.2 Bior 5.5 Symlets 8 Coiflets 4 Rev. Bior 6.8 All
Average PSNR 6.094 5.832 5.938 6.010 5.924 5.930 5.935 5.943 5.95
Minimum PSNR 6.056 5.796 5.933 5.970 5.912 5.924 5.931 5.937 5.932
591
Authorized licensed use limited to: CHONGQING UNIVERSITY. Downloaded on March 21, 2009 at 02:00 from IEEE Xplore. Restrictions apply.
Maximum PSNR 6.109 5.895 5.944 6.040 5.932 5.939 5.939 5.951 5.968
Table 5. Watermark Extraction with Noise Salt and Speckle Gaussian Pepper Average Average Average Wavelet PSNR PSNR PSNR Haar 6.074 6.077 6.066 Daubechies 5.858 5.859 5.866 Daubechies 32 5.938 5.938 5.938 Bior 2.2 5.988 5.988 5.983 Bior 5.5 5.929 5.929 5.930 Symlets 8 5.933 5.934 5.934 Coiflets 4 5.936 5.936 5.936 Rev. Bior 6.8 5.942 5.942 5.941
[3]
[4]
[5]
[6] [7] [8]
6. Conclusions Our study shows that watermarking with the Haar wavelet outperforms the other watermarking transforms tested, with 2 dB for embedding watermarks (Table I), 6-7 dB for embedding with noise (Table II), and is slightly better (about 3%) than other transforms tested for watermark extraction (examples in Table III, averages in Table IV) and watermark extraction with noise (Table V). The Discrete Wavelet Transform has historically shown its suitability for watermarking applications. It effectively allows the embedding of a watermark at higher level frequencies, which are not as visible to the human eye, via the access to the wavelet coefficients in the HL and LH detail sub-bands. However, not much attention has been given to which wavelet may be preferred. For both the impact on the original image and for the recovery of the embedded watermark, the Haar wavelet, both visually and objectively by measured by PSNR, outperforms the other families tested. This remains true when three types of noise are added, including Gaussian, speckled and salt and pepper, as well as when the watermarked image remains uncorrupted. In almost every situation the Haar wavelet repeatedly outperforms the others. Therefore, the size, type and complexity of the image, and the introduction of noise does not seem to change the advantage that the simple, but effective, Haar Wavelet displays for this watermarking application.
[9] [10]
[11]
[12]
[13]
[14]
[15]
techniques,“ Proceedings of the IEEE, Volume 87, Issue 7, 1999, pages 1079-1107. Darko Kirovski and Fabian A. P. Petitcolas, “Blind Pattern Matching Attack on Watermarking Systems,” IEEE Transactions on Signal Processing, Volume 51, Number 4, 2003, pages 1045-1053. F. A. P. Petitcolas, “Watermarking Schemes Evaluation” IEEE Signal Processing Magazine, Volume 17, September 2000. pages 58-64. F. Hartung and M. Kutter, Stefan Katzenbeisser and Fabien A. P. Petitcolas, editors, Information Hiding Techniques for Steganography and Digital Watermarking, Artech House, 2000. J. Weiner and K. Mirkes, Watermarking, Inst. Paper Chemistry, Appleton, WI, 1972. Juergen Seitz, Digital Watermarking for Digital Media, Information Science Publishing, 2005. Neil F. Johnson, Zoran Duric, and Sushil Jajodia, Information Hiding : Steganography and Watermarking Attacks and Countermeasures (Advances in Information Security, Volume 1) (Advances in Information Security), Kluwer Academic Publishers, Norwell, MA, 2006. Michael Weeks, Digital Signal Processing Using MATLAB and Wavelets, Infinity Science Press, 2006. G. Langelaar, I. Setyawan, and R. L. Lagendijk, “Watermarking Digital Image and Video Data,” IEEE Signal Processing Magazine, Number 17, September 2000, pages 20-43. MATLAB Documentation, Image Processing Toolbox User’s Guide, Release 14, The MathWorks, Inc. 3 Apple Hill Drive, Natick, MA 01760-2098, 2006. H. Inoue, A. Miyazaki and T. Katsura, “An image watermarking method based on the wavelet transform,” ICIP 99. Proceedings. 1999 International Conference on Image Processing, (1), 1999, pages 296-300. Stephane Jaffard, Yves Meyer and Robert D. Ryan, Wavelets Tools for Science and Technology, Society for Industrial and Applied Mathematics (SIAM), 2001. Stephane Mallat, “A Theory for Multiresolution Signal Decomposition: The Wavelet Representation,” IEEE Pattern Analysis and Machine Intelligence, Volume 11, Number 7, 1989, pages 674-693. Some test images are from: E.R. Vrscay, F. Mendivil, H. Kunze, D. La Torre, S.K. Alexander, and Bruno Forte, “Waterloo Repertoire GreySet (1 and 2)”, Waterloo Fractal Coding and Analysis Group website, Accessed October 2007, University of Waterloo, Waterloo, Ontario, Canada, http://links.uwaterloo.ca/greyset1.base.html and http://links.uwaterloo.ca/greyset2.base.html.
References [1] F. A. P. Petitcolas, R.J. Anderson, R. J. and M. G. Kuhn, “Information hiding - A survey,” Proceedings of the IEEE, Volume 87, Issue 7, 1999, pages 1062-1078. [2] F. Hartung and M. Kutter, “Multimedia watermarking
592
Authorized licensed use limited to: CHONGQING UNIVERSITY. Downloaded on March 21, 2009 at 02:00 from IEEE Xplore. Restrictions apply.