Beruflich Dokumente
Kultur Dokumente
Abstract—A semifragile watermarking scheme for remote Some semi-fragile watermarking schemes of multiple
sensing images is presented. The suggested scheme uses the band images also consider only one band or process each
signature of the multispectral or hyperspectral image to embed band separately [3], [4]. Note that, if the bands are marked
the mark and detects a forgery of the marked image, e.g. a
tampered region. The original image is segmented in three- separately, the changes in the signature curves can be
dimensional blocks and, for each block, a discrete Wavelet uneven. Hence, the shapes of the signatures may vary, which
transform (DWT) and a tree structured vector quantizer is may lead to a misclassification of the image. Because of
built. These trees are manipulated using an iterative algorithm this, a method which preserves the shapes of the signatures
until the resulting image satisfies some selected conditions. is highly demanded.
Each tree is partially modified accordingly to a secret key
in order to avoid tampering attacks. This key determines the There are different previous works dealing with satellite
internal structure of the tree and, also, the resulting distortion. image watermarking [5], [6], [7], [8], [9], [10], [11]. Ho
In order to make the marked image robust against near- et al. [5] uses a satellite image and decomposes it into two
lossless compression, the trees are built using only the LL mutually orthogonal sub-fields, but only uses one band of the
sub-band of the DWT. The results show that the method satellite image and only one field for watermarking purposes.
works correctly with remote sensing images and detects copy-
and-replace attacks from segments of the same image, whilst Qin et al. [6] presents a semi-fragile watermarking scheme
allowing for JPEG2000 compression. based on wavelet transforms. The edge and texture of the
remote sensing image are extracted and the watermark is
Keywords-Forensic Watermarking, tampering detection, re-
mote sensing images embedded only in the edge character. In this case, once
again, only one band is marked. Wang et al. [7] presents
a watermarking scheme to preserve a digital content, but
I. I NTRODUCTION
only uses one band of the hyperspectral image of the Indian
To prevent the alteration of digital files and detect ma- Remote Sensing System. Sal and Graña [8] describes an
nipulations, watermarking schemes have been developed evolutionary algorithm which marks an image based on
in the last few years. The embedded watermarks must be the manipulation of the discrete cosine transform (DCT)
imperceptible and can used to determine the integrity of the computed for each band of the image. Tamhankar et al. [9]
digital files. describes a method to embed one mark into the hyperspectral
Some watermarking schemes [1], [2] insert a watermark image using the whole signature, but it does not allow
into an image, such that the watermark is modified or compression of the hyperspectral image. Finally, Serra et al.
destroyed when the marked image is manipulated. In this [10] and Serra and Megı́as [11] use the whole signature to
authentication process, two different approaches, namely embed a watermark, and Serra and Megı́as [11] implements
fragile watermarking and semi-fragile watermarking, can be an LSB extraction in each pixel value, making the scheme
used. In fragile schemes, all modifications are detected as robust against near-lossless compression.
tampering. Therefore any kind lossy compression or filters In this paper, a semi-fragile watermarking scheme specifi-
can not be applied to the marked image without removing cally developed for remote sensing images is presented. The
the embedded mark. On the other hand, semi-fragile schemes method can be tuned to embed the mark according to band
allow some degree of compression and small modifications relevance, depending on the content and the signatures (also
(or “attacks”) of the marked images. This makes it possible, known as the spectral reflectance curve) to be protected.
for example, to create a compressed version of the image The suggested method applies the discrete Wavelet transform
which can be distributed electronically (possibly with a (DWT) and a vector quantization approach for a group of
reduced price) but maintaining the original watermark. Both selected bands. The image is segmented in three-dimensional
fragile and semi-fragile watermarking schemes can be used blocks of a given size which determines the spatial resolution
for tampering detection and localization. of the embedding and detection algorithm. For each block,
RED
GRN
Near Infrared Middle Infrared
Reflected Infrared
and the obtained LL sub-bands are used to build a Tree- 60
Structured Vector Quantization (TSVQ) tree. This TSVQ Vegetation Dry soil
Reflectance(%)
(5% water)
tree is then modified using an iterative algorithm until it
satisfies some criterion. Once the target value (generated 40
Wet soil
with a secret key) is reached, the marked block is obtained (20% water)
produces a different criterion for each block in order to avoid Turbid river water
0
copy-and-replace attacks. The use of the LL DWT sub-band 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4
for each spectral band makes it possible to obtain robustness Wavelength (micrometers)
against near-lossless compression attacks and, at the same Figure 1: Examples of signatures.
time, strong modifications are detected as tampering.
This paper is organized as follows. In Section 2, back-
ground about watermarking, remote sensing images, lossy distortion in the marked image and a trade-off between
compression and vector quantization is provided. Section 3, robustness and perceptual quality must be achieved.
describes the watermarking strategy and the mark embed- In remote sensing imaging applications, the most useful
ding and mark retrieval processes. Section 4 presents the schemes aimed to detect changes in the image are semi-
results obtained with the suggested scheme and, finally, the fragile watermarking systems.
conclusions of this work are drawn in Section 5.
B. Remote sensing images
II. BACKGROUND
Remote sensing images contain information about an area
In this section, the basic concepts used in the method of the surface of the Earth. Each pixel is represented by an
presented in this paper, namely watermarking, remote sens- endmember, which is a set of values obtained for different
ing images, lossy compression and vector quantization, are frequencies of the light spectrum (bands) obtained by a
shortly introduced. remote sensor. The signature of each pixel of the image is
related to the different materials in that area, such as water,
A. Watermarking forest or minerals. Fig. 1 shows the different signatures for
Watermarking consists of imperceptibly embedding some a light reflectance for clear lake water, turbid river water,
information into a cover object (e.g. a remote sensing vegetation, dry soil and wet soil [12].
image) to produce a marked version of the same object. Two of the most relevant problems with these images is
The watermarking process alters the original data file by their huge size and the cost to obtain them. A typical hyper-
modifying the content in order to embed the mark. Most spectral image covering a small region of a few kilometres
schemes require a common (secret) key both at the embedder contains millions of pixels. As an example, the Airbone
at the detector which is usually introduced to endorse the Visible / Infrared Imaging Spectrometer (AVIRIS) images
system with security features. The detection process needs contain 224 bands and, usually, 16 bits are used for the
the (secret) key to determine whether the mark is or not values in each band. Images with lower resolutions are often
embedded in the test object. With a blind detection method, referred to as “multispectral”. This is the case of Landsat
no further information is needed to detect the mark (i.e. the images, which use 8 bands for each pixel and 8 bits for the
original unmarked content is not required). values of each band.
Semi-fragile schemes are able to overcome some minor
modifications, as those produced by near-lossless compres- C. Lossy compression of remote sensing images
sion, but reveal the existence of strong manipulations (also Lossy compression methods remove information which
referred to as “attacks” in the watermarking literature), such is not significant for image reconstruction. The amount of
as copy-and-replace, or excessive information removal by removed information should depend on the user purposes.
means of cropping or lossy compression. Semi-fragile wa- Some experiments [13] carried out with remote sensing
termarking makes unnecessary to mark different compressed images, show that it is possible to achieve relatively high
versions of the same image independently and, thus, reduces compression ratios without removing critical information.
the cost required to distribute the same marked image with The general coding scheme must be adapted to the particular
different degrees of quality. characteristics of the source, in order to maximize both the
On the other hand, robust watermarking allows stronger compression ratio and the image fidelity. In this case, the
types of manipulations, including compression, filtering or most important issue is that the remote sensing images are
geometrical attacks. Usually, robustness requires significant 3D, where two dimensions are spatial and one is spectral.
332
Bands Selection Block Construction
Vector Construction
16 bands
LL LH
32
32x32
512
64
16
HL HH
16
32
16
LL Block
64 TSVQ vectors
512 Wavelet
Hyperspectral image 512x512 reduced image
333
Key PRNG
image into the marked image, a pseudo-random sequence
Block
is chosen to determine the values used as criteria to select TSVQ no
forged
the compression subtree in the pruning algorithm. Thus, Marked DWT vectors Criterion
it is much more difficult to find a pattern revealing the block established
??
334
Table I: PSNR of the marked image for each band.
Band PSNR PMP Adi Band PSNR PMP Adi Band PSNR PMP Adi
1 61.15 32.33 18.02 6 62.39 32.10 15.19 11 62.76 32.18 15.14
2 61.29 32.10 16.68 7 62.48 32.17 15.08 12 61.24 32.40 18.40
3 61.82 32.19 16.37 8 62.18 32.18 15.82 13 61.71 32.19 17.05
4 61.94 32.19 16.34 9 58.07 32.56 25.94 14 60.16 32.27 19.12
5 62.26 32.12 15.40 10 62.86 32.05 14.89 15 60.74 32.32 18.56
16 58.92 32.53 23.48
Figure 5: Original (a), marked (b) and tampered (c) images and tamper location (d) for band 2.
by the embedding process, whereas the other 97.70% of the and 5d show a tampered image and the tamper location
pixels are identical to the original. as detected by the proposed scheme. A forest region was
The PSNR for each of the modified bands is shown in copied and replaced by a mineral region, next to the forest.
Table I, where “PMP” stands for “Percentage of Modified This forgery has been perfectly detected by the suggested
Pixels” and “Adi” stands for “Average difference” of the scheme. In addition, no false positive forgeries have been
modified pixels. Notice that the PSNR is quite high. On detected in the performed experiments.
average, the PSNR for the different modified bands is 61.37 Finally, the robustness of the scheme has been checked
dB. The number of modified pixels (PMP) is 32.24% with against compression attacks with the JPEG2000 algorithm.
an average difference of 17.59 (where the range of pixel Table II shows some compression attacks, in bits per pixel
values is from 1, 200 to 12, 500). (bpp), and the robustness of the scheme against them. Note
A comparison of different watermarking schemes for that the mark survives always, and therefore the scheme is
hyperspectral images is provided in [11]. Due to space robust against JPEG2000 compression.
limitations, such a comparative analysis (with the methods
introduced in Section I) is not included in this paper. How- V. C ONCLUSIONS
ever, the method presented here provides similar distortion A semi-fragile watermarking method for remote sensing
results compared to that of [11] (which overcomes the images based on TSVQ and DWT is presented. The method
results of the other schemes) and better robustness against uses the information in a set of bands at the same time,
JPEG2000 compression. and thus, it takes advantage of both spatial and spectral
redundancy for marking purposes. Basically, the original
Table II: Robustness against JPEG2000 compression attacks. image is segmented in three dimensional blocks and a tree
Compr. Mark Max structured vector quantizer is built for the LL DWT sub-
Adi PMP
(bpp) survival diff. band of each block. The original LL sub-band of each block
8 yes 1.12 5 50.45% is replaced by a new one generated by substituting each
7 yes 1.56 10 71.86% original vector by the closest centroid in the selected subtree.
6 yes 2.39 17 83.81%
5 yes 4.14 32 91.26% This process is repeated until a certain stopping criterion
is satisfied. Each block generates a different subtree and a
secret key is used to avoid copy-and-replace attacks between
A random block selection and replacement into the image blocks. The results show that copy-and-replace attack of a
has been performed into the marked image. Figures 5c region of the image is detected by the watermarking scheme.
335
The detection process is a simple one, since a tree is built [8] D. Sal and M. Graña, Studies in Computational Intelligence.
for each block and the selected tree property is tested. Springer Berlin / Heidelberg, 2008, vol. 133/2008, ch. A Mul-
If a block satisfies such a property, then it has not been tiobjective Evolutionary Algorithm for Hyperspectral Image
Watermarking, pp. 63–78.
forged, otherwise the detection process reports tampering.
In addition, the use of the DWT in each band provides [9] H. Tamhankar, L. Bruce, and N. Younan, “Watermarking of
robustness against JPEG2000 compression attacks. hyperspectral data,” Geoscience and Remote Sensing Sympo-
Future research lines in this subject include the study sium, 2003. IGARSS ’03. Proceedings. 2003 IEEE Interna-
tional, vol. 6, pp. 3574–3576 vol.6, July 2003.
of optimal subtree selection criteria for pruning purposes,
the use of different Wavelet transforms and their effect on [10] J. Serra, D. Megias, J.Herrera-Joancomartı́, and J. Minguillón,
distortion and robustness against compression and the use “Multiband semigrafile watermarking for multi and hyper-
of smaller block sizes to increase the resolution of the spectral images based on iterative tree structured vector quan-
tampering detection. The possibility of using different sizes tization,” in Proceedings of Image and signal processing for
remote sensing XII, L. Bruzzone, Ed., vol. 6365. Stockholm:
for different areas of the image can also be addressed. It
SPIE, September 2006, pp. 63 650O1–63 650O10.
would be advisable to use smaller blocks in the region of
interest (ROI) and bigger blocks in other areas. [11] J. Serra-Ruiz and D. Megı́as, “A novel semifragile forensic
watermarking scheme for remote sensing images,” Interna-
ACKNOWLEDGEMENTS tional Journal of Remote Sensing, 2010, in press.
This work is partially supported by the Spanish Ministry of [12] R. B. Smith, “Introduction to hyperspec-
Science and Innovation and the FEDER funds under the grants tral imaging,” Aug 2006, available at
TSI2007-65406-C03-03 E-AEGIS and CONSOLIDER-INGENIO http://www.microimages.com/getstart/hyprspec.htm.
2010 CSD2007-00004 ARES. [13] J. Minguillón, J. Pujol, J. Serra, and I. Ortuño, “Influence of
lossy compression on hyperspectral image classification,” in
R EFERENCES Proceedings of Data Mining’2000, Cambridge, UK, July 5-7
2000, pp. 545–554.
[1] M. Yeung and F. Mintzer, “An invisible watermarking tech-
nique for image verification,” in International Conference on [14] A. Gersho and R. M. Gray, Vector Quantization and Signal
Image Processing, IEEE, vol. 2, Santa Barbara, CA, Oct. Compression, ser. Communications and Information Theory.
1997, pp. 680–683. Norwell, MA, USA: Kluwer Academic Publishers, 1992.
[2] J. Fridrich, “Security of fragile authentication watermarks [15] L. Breiman, J. H. Friedman, R. A. Olshen, and C. J. Stone,
with localization,” in Security and Watermarking of Multi- Classification and Regression Trees. Wadsworth Interna-
media Contents, SPIE, Ed., vol. 4675. San Jose, CA: SPIE, tional Group, 1984.
January 2002, pp. 691–700.
336