WATERMARKING FOR IMAGE AUTHENTICATION
Min Wu Bede Liu
Department of Electrical Engineering Princeton University, Princeton, NJ 08544, USA Fax: +1-609-258-3745
fminwu, liu
g@ee.princeton.edu ABSTRACT
A data embedding method is proposed for image authentication based on table look-up in frequency do- main. A visually meaningful watermark and a set of simple features are embedded invisibly in the marked image, which can be stored in the compressed form.
The scheme can detect and localize alterations of the original image, such as the tempering of images ex- ported from a digital camera.
Keywords:
digital watermark, image authentica- tion.1. INTRODUCTION
Digital information revolution has brought about many advantages and new issues. With the ease of editing and perfect reproduction, the protection of ownership and the prevention of unauthorized manipulation of digital audio, image, and video materials become im- portant concerns. Digital watermarking, a scheme to embed special labels in digital sources, has made con- siderable progress in recent years. There are several categories of watermarking schemes [1]. Among them, fragile watermarking is a technique to insert a signa- ture for image authentication. The signature will be altered when the host image is manipulated. In this paper, we focus on watermarking for digital image au- thentication.
An e ective authentication scheme should have the following desirable features:
1. to be able to determine whether an image has been altered or not;
2. to be able to locate any alteration made on the image;
Thisresearchissupp ortedbyaNewJerseyStateR&DEx-
3. to be able to integrate authentication data with host image rather than as a separate data le;
4. the embedded authentication data be invisible under normal viewing conditions;
5. to allowthe watermarked imagebe stored in lossy- compression format.
Previously published methods for image authenti- cation do not satisfy all the requirements. The digital signature proposed in[2], as well as the content based signature reported in [3] and [4] do not satisfy require- ments 2 and 3. The pixel-domain scheme [5] can not be stored in lossy compression format. In addition, two frequency domain data hiding schemes [6][7] may be used for authentication, but they cannot always locate alteration and the distortion introduced by embedding is larger than the proposed scheme.
This paper presents a new authentication scheme by embedding a visually meaningful watermark and a set of simple features in the frequency domain of an image via table look-up. This scheme can be applied to compressed image using JPEG or other techniques such as Wavelet compression, and the marked image can be kept in the compressed format. Any alteration made on the marked image can be localized, making it suitable for \trustworthy" digital camera. Furthermore, this scheme is computationally ecient, and can be applied to video.
2. TWO ASPECTS OF PROPOSED AUTHENTICATION APPROACH
We shall present our approach in the context of grey scale JPEG compressed images. The extension to im- ages compressed using other means and to color im- ages are reasonably straightforward and discussed in Section 7.
A block diagram for the embedding of data is given in Fig. 1 which, aside from the block labeled \embed",
is identical to that of the JPEG compression [8]. Wa- termark is inserted into the quantized DCT coecients via a look-up table.
Marked Coeff.
Embed block
DCT Quant.
Coding Entropy (LUT)
Data to be embedded
JPEG Image Marked Lookup Table
Image Original
Figure 1: Block Diagram of Embedding Process In the next section, we present how to embed data.
The question of what data are used for authentication is discussed in Section 4.
3. PROPOSED FREQUENCY-DOMAIN DATA EMBEDDING SCHEME
3.1. Embedding Binary Bits Via Table Look-up
A proprietary look-up table (LUT) is generated before- hand by the owner of the image or the digital camera manufacturers. The table maps every possible value of JPEG coecient randomly to \1" or \0" with the con- straint that runs of \1" and \0" are limited in length.
To embed a \1" in a coecient, the coecient is un- changed if the entry of the table corresponding to that coecient is also a \1". If the entry of the table is a \0", then the coecient is changed to its nearest neighbor- ing values for which the entry is \1". The embedding of a \0" is similar. This process can be abstracted into the following formula whereviis the original coecient,vi0 is the marked one,bi is the bit to be embedded in, and
LUT() is the mapping by look-up table:
v
i 0
=
8
>
<
>
: v
i ifLUT(vi) =bi
v
i
+
ifLUT(vi)6=bi, and= minjdjfd2Z:LUT(vi+d) =big The extraction of the signature is simply by looking up the table. That is,
^
b
i
=
LUT(
vi0)
where ^ is the extracted bit.
The basic idea of the embedding process is also il- lustrated by the example in Fig. 2. In this example, zeros are to be embedded in two AC coecients with values \-73" and \-24" of an 8x8 image block. The en- try in the table for coecient value \-73" is \1". So to embed a \0", we go to its closest neighbor for which the entry is \0". In this case, \-73" is changed to \-74".
Since the entry for coecient value \24" is \0", it is unchanged.
24 0 24
-73
24 . . . .23
1 . . . 0 0
-74 24
0 Embed "0" (changed)
-74
Embed "0" (unchanged)
-74 -73 0
-72
1 1
25 -75
1
36 -73 8 3 1 1 0
24 5 -10 -2 1 0 0 0 2 -1 0 -1 -1
0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 5
7 0
0
1
0
X
embed
Original Marked
. . . . . . . . . . . .
. . . . Look-up
(LUT) Table
Figure 2: Frequency-domain Embedding Scheme
3.2. Considerations For Minimizing Distortion
Several steps are taken to ensure that the markings are invisible:
The run of \1" and \0" entries in the LUT is con- strained to avoid excessive modi cation on the coecients;
DC coecient in each block is not changed to avoid blocky e ect unless the quantization step is very small;
Smallvalued coecients are not modi ed to avoid high frequency distortion.
The above constraints make the number of bits that can be embedded varies signi cantly fromblock to block.
Also, extraction error may occur due to image format conversion and other causes. For these reasons, we choose to embed one bit per 8x8 block by embedding the same bit to all embeddable coecients in the block and detecting that bit by majority voting. The detec- tion of an error in the embedded bit signals alteration of that block, which will be discussed in next section.
Better encoding schemes will be discussed in Section 7.
The constraint of leaving DC coecients and small valued AC coecients unchanged makes it impossible
to embed data in smooth blocks where all AC coe- cients are very small hence unembeddable. We shall discuss this problem in Section 5.
4. CHOICE OF EMBEDDED DATA FOR AUTHENTICATION
The authentication data we embed in an image consists of a visually meaningful binary pattern and some con- tent features. The visually meaningful pattern serves as a quick check for alteration and location of possible alteration. A simple example of content features is the most signi cant bit of macroblock mean intensity. An- other example is the sign of intensity di erence between macroblocks. The combination of these two types data is suitable for such applications as imageauthentication for \trustworthy" digital cameras. The embedding and verifying procedures are summarized in Fig. 3.
Compressed Marked
Image Original
Image
block
DCT Quant. Embed
Features PatternBinary
Entropy Coding Marked
Coeff.
(LUT) Lookup Table
block
DCT Quant. Extract PatternBinary Embedded
Features
Visualize Compare
Features (LUT) Lookup Table Test Image
(decompressed)
Figure 3: Embedding and Extraction Procedures
5. SMOOTH REGION EMBEDDING
Authentication data cannot be embedded in smooth blocks because the AC coecients are small. In a typi- cal natural image such as the one shown in Fig. 4, about 20% of the blocks are smooth, which is illustrated in Fig. 5.
We propose a solution to embed data in smooth re- gion by embedding data in blocks that are not smooth but whose location bears a xed relationship to the smooth block in question. A simple implementation, called backup embedding, is illustrated in Fig. 6. In- stead of embedding one bit independently in each block,
Figure 4: Original Image ( 75% JPEG Compression ) we take the macroblock as a unit. We use two bits to embed data corresponding to the current macroblock, and use the other two bits to embed a backup copy of data corresponding to the companion macroblock. In the illustration of Fig. 6, the companion blocks are half picture height apart.
Figure 5: Smooth Blocks of Fig. 4 (shown in black)
6. EXPERIMENTAL RESULTS
A JPEG compressed image of size 640x432 is shown in Fig. 4. Shown in Fig. 7(a) is the same image but with a 40x27 binary pattern \PUEE" of Fig. 7(b) and the MSB of macroblock mean intensity embedded in.
This image is indistinguishable from the original. The smooth regions of the image is shown in Fig. 5. For these blocks, the backup embedding is used.
The marked image is modi ed by changing \Prince- ton University" to \Alexander Hall" and \(c) Copy- right 1997" to \January 1998", shown in Fig. 8(a). This
000000 111111 000000 111111 H
1 1 0 1
1 1 1 0 16 x 16
(i, j)
(i, j+H/2) 4 bits
backup backup
Figure 6: Backup Embedding For Smooth Region image is again stored in the JPEG format. For ease of visualization, we shall denote the block that embeds
\0" as a black dot, the block that embeds \1" as a white dot, and the block with no majorityas a gray dot.
With these notation, Fig. 8(b) and Fig. 8(c) show the extracted binary pattern and content feature match- ing result of the edited image. The random pattern corresponding to the altered blocks are clearly identi - able. Also, very few unembeddable data are shown as isolated gray dots by the backup scheme.
Figure 7: ab (a) Marked Image; (b) Embedded Binary Pattern.
Figure 8: bjca (a) Marked Image After Editing (stored in 75% JPEG); (b) Extracted Binary Pattern from Edited Image; (c) Feature Matching Result.
7. IMPROVEMENT AND EXTENSIONS 7.1. Double Watermarking For Authentication and Ownership Veri cation
Previous work on double watermarking mainly empha- sized on embedding multiple labels (such as ownership label, recipient label, etc.) using the same embedding approach. No published work discussed to verify own- ership as well as signal and locate alteration. A key issue for this kind of double watermarking is that no original \true" image can be used since the validity of this original image has not been veri ed yet.
The proposed new scheme can be combined with a previous robust watermarking and detection scheme on ownership veri cation [9] for double watermarking.
This double watermarking has been tested successfully.
7.2. Embedding More Data
We have described previously a method to embed 1 bit per 8x8 block. The amount of data embedded can be increased by more sophisticated backup and coding.
7.3. Multi-level Data Embedding and Unequal Error Protection
We mentioned in Section 4 that two sets of data, namely, a meaningful visual pattern and a set of low-level con-
tent features, are embedded in the image for authen- tication purpose. Multi-level data embedding and un- equal error protection are possible in order to embed several sets of data with di erent levels of error protec- tion.
7.4. Other Compression Format
Besides JPEG compressed image,we also tested Wavelet compression and found our approach e ective.
7.5. Color Image and Video
For color images, we may work in YCrCb coordinates and use the new approach to mark luminance compo- nents while leaving chrominance components unchanged;
or we may apply the approach to chrominance compo- nents as well to embed more data. We may also work in other color coordinates, such as RGB.
The proposed scheme can be applied to MPEG com- pressed digital video. A simple approach would be to mark I-frames of video streams using our new scheme.
In addition, I-frame serial number can be used as part of embedded data to detect modi cation such as frame reordering and frame dropping.
8. CONCLUSION
In this paper, we have presented a frequency domain technique for image authentication in which the sig- nature is embedded in the image and the marked im- age can be kept in the compressed form. The scheme can detect tempering of the marked image and can locate where the tempering has occurred. Potential applications include trust-worthy digital cameras and camcorders. Image samples and other information are availableat http://www.ee.princeton.edu/minwu/rsch authwmark.html . Refer to the les attached in the CD-ROM for full-size high-resolution images of Fig. 4, 7(a), and 8(a).
9. REFERENCES
[1] F. Mintzer, G. W. Braudaway, M. M. Yeung: \Ef- fective and Ine ective Digital Watermarks",ICIP, 1997
[2] G. L. Friedman: \The Trustworthy Digital Cam- era: Restoring Credibility to the Photographic Im- age",IEEE Trans. on Consumer Electronics, Nov.
1993.
[3] M. Schneider, S-F. Chang: \A Robust Content Based Digital Signature for Image Authentica- tion",ICIP, 1996.
[4] D. Storck: \A New Approach to Integrity of Dig- ital Images",IFIP Conf. on Mobile Communica- tion, 1996.
[5] M. M. Yeung, F. Mintzer: \An Invisible Water- marking Technique for Image Veri cation",ICIP, 1997.
[6] M. D. Swanson, B. Zhu, A. H. Tew k: \Robust Data Hiding for Images", IEEE DSP Workshop, 1996.
[7] E. Koch, J. Zhao: \Towards Robust and Hidden Image Copyright Labeling", IEEE Workshop on Nonlinear Signal and Image Processing, 1995.
[8] G. K. Wallace: \The JPEG Still Picture Compres- sion Standard",IEEE Trans. on Consumer Elec- tronics, vol.38, no.1, pp18-34, 1992
[9] W. Zeng, B. Liu: \On Resolving Rightful Owner- ships of Digital Images by Invisible Watermarks", ICIP, 1997.
