Exposing_Digital_Forgeries_from_JPEG_Ghosts.pdf

154

IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 4, NO. 1, MARCH 2009

IV. C ONCLUSION

The embedding efciency of codes from the ZZW embedding con-

struction [11] follows the upper bound on embedding efciency. The

distance to the bound in the zero-payload limit can be expressed in a

closed form using the code parameters. The limit could be used to order

codes by their asymptotic performance. We note that the embedding

construction for embedding also proposed in [11] approaches the

bound on embedding efciency of ternary codes with the same limit

(6). This is because the ternary bound increases by 1 compared to the

binary bound (as explained in Section I) and, as shown in [11], the em-

bedding efciency of ZZW code families is also larger by 1.

Exposing Digital Forgeries From JPEG Ghosts

Hany Farid

Abstract— When creating a digital forgery, it is often necessary to com-

bine several images, for example, when compositing one person’s head onto

another person’s body. If these images were originally of different JPEG

compression quality, then the digital composite may contain a trace of the

original compression qualities. To this end, we describe a technique to de-

tect whether the part of an image was initially compressed at a lower quality

than the rest of the image. This approach is applicable to images of high and

low quality as well as resolution.

Index Terms— Digital forensics, digital tampering.

R EFERENCES

[1] J. Bierbrauer and J. Fridrich, “Constructing good covering codes for

applications in steganography,” Lect. Notes Comput. Sci. Trans. Data

Hiding and Multimedia Security , vol. 4920, pp. 1–22, 2008.

[2] C. Cachin, D. Aucsmith, Ed., “An information-theoretic model for

steganography,” in Proc. 2nd Int. Workshop Information Hiding ,New

York, Apr. 14–17, 1998, vol. 1525, pp. 306–318.

[3] R. Crandall, Some notes on steganography. Steganography Mailing

List, 1998. [Online]. Available: http://os.inf.tu-dresden.de/west-

feld/crandall.pdf.

[4] T. Filler and J. Fridrich, “Binary quantization using belief propaga-

tion over factor graphs of LDGM codes,” presented at the 45th Annu.

Allerton Conf. Communication, Control, Computing Allerton, IL, Sep.

26–28, 2007.

[5] J. Fridrich, P. Lisonek, and D. Soukal, J. L. Camenisch, C. S. Collberg,

N. F. Johnson, and P. Sallee, Eds., “On steganographic embedding ef-

ciency,” in Proc. 8th Int. Workshop Information Hiding , New York,

Jul. 10–12, 2006, vol. 4437, pp. 282–296.

[6] J. Fridrich and T. Filler, E. J. Delp and P. W. Wong, Eds., “Practical

methods for minimizing embedding impact in steganography,” in Proc.

SPIE, Electronic Imaging, Security, Steganography, Watermarking of

Multimedia Contents IX , San Jose, CA, Feb. 1, 2007, vol. 6505, pp.

02–03.

[7] J. Fridrich, M. Goljan, and D. Soukal, M. Barni, J. Herrera, S. Katzen-

beisser, and F. Pérez-González, Eds., “Efcient wet paper codes,” in

Proc. 7th Int. Workshop Information Hiding , Barcelona, Spain, Jun.

6–8, 2005, pp. 204–218.

[8] F. Galand and G. Kabatiansky, “Information hiding by coverings,” in

Proc. IEEE Information Theory Workshop , Paris, France, Apr. 4, 2003,

pp. 151–154.

[9] A. D. Ker, J. L. Camenisch, C. S. Collberg, N. F. Johnson, and P. Sallee,

Eds., “Batch steganography and pooled steganalysis,” in Proc. 8th Int.

Workshop Information Hiding , New York, Jul. 10–12, 2006, vol. 4437,

pp. 265–281.

[10] J. Kodovský, J. Fridrich, and T. Pevný, J. Dittmann and J. Fridrich,

Eds., “Statistically undetectable JPEG steganography: Dead ends,

challenges, and opportunities,” in Proc. 9th ACM Multimedia Security

Workshop , Dallas, TX, Sep. 20–21, 2007, pp. 3–14.

[11] W. Zhang, X. Zhang, and S. Wang, K. Solanki, K. Sullivan, and U.

Madhow, Eds., “Maximizing steganographic embedding efciency by

combining Hamming codes and wet paper codes,” in Proc. 10th Int.

Workshop Information Hiding , New York, Jun. 19–21, 2008, pp. 60–71.

I. I NTRODUCTION

Recent advances in digital forensics have given rise to many tech-

niques for detecting photographic tampering. These include techniques

for detecting cloning [1], [2]; splicing [3]; resampling artifacts [4],

[5]; color lter-array aberrations [6]; disturbances of a camera’s sensor

noise pattern [7]; chromatic aberrations [8]; and lighting inconsisten-

cies [9]–[11]. Although highly effective in some situations, many of

these techniques are only applicable to relatively high-quality images.

A forensic analyst, however, is often confronted with low-quality im-

ages in terms of resolution and/or compression. As such, there is a need

for forensic tools that are specically applicable to detect tampering in

low-quality images. This is particularly challenging since low-quality

images often destroy any statistical artifacts that could be used to detect

tampering.

Along these lines, Ye, et al. developed a technique to estimate the

local JPEG compression blocking artifacts [12]—inconsistencies in

these artifacts were then used as evidence of tampering. Luo et al.

developed a technique to detect inconsistencies in JPEG blocking

artifacts that arise from misalignments of JPEG blocks relative to their

original lattice [13]. And He et al. developed a technique to detect

local traces of double JPEG compression [14] (this correspondence

expands on a global approach to detect double compression [15]).

A complementary approach to detect tampering in low-quality im-

ages is presented here. This approach detects tampering which results

when part of a JPEG image is inserted into another higher quality JPEG

image, for example, when one person’s head is spliced onto another

person’s body, or when two separately photographed people are com-

bined into a single composite. This approach works by explicitly deter-

mining if part of an image was originally compressed at a lower quality

relative to the rest of the image.

Manuscript received February 27, 2008; revised October 08, 2008. First pub-

lished February 3, 2009; current version published February 11, 2009. This work

was supported in part by Adobe Systems, Inc., in part Microsoft, Inc., in part

by the National Science Foundation (CNS-0708209), in part by the U.S. Air

Force Grant (FA8750-06-C-0011), and in part by the Institute for Security Tech-

nology Studies at Dartmouth College under Grants from the Bureau of Justice

Assistance (2005-DD-BX-1091) and the U.S. Department of Homeland Secu-

rity (2006-CS-001-000001). The points of view or opinions in this document are

those of the author and do not represent the ofcial position or policies of the

U.S. Department of Justice, the U.S. Department of Homeland Security, or any

other sponsor.The associate editor coordinating the review of this manuscript

and approving it for publication was Dr. M. Kivanc Mihcak.

The author is with the Department of Computer Science at Dartmouth Col-

lege, Hanover, NH 03755 USA (e-mail: farid@cs.dartmouth.edu).

Color versions of one or more of the gures in this paper are available online

at http://ieeexplore.ieee.org.

Digital Object Identier 10.1109/TIFS.2008.2012215

IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 4, NO. 1, MARCH 2009

155

Fig. 1. Shown in panel (a) is the sum of squared differences between coefcients quantized by an amount , followed by a second quantization in the range

(horizontal axis)—this difference reaches a minimum at . Shown in panel (b) is the sum of squared differences between coefcients

quantized initially by an amount followed by , followed by quantization in the range (horizontal axis)—this difference reaches a

minimum at and a local minimum at , revealing the original quantization.

Fig. 2. Shown in the top left panel is the original image from which a central 200 200 region was extracted, saved at JPEG quality 65, and reinserted into the

image whose original quality was 85. Shown in each subsequent panel is the difference between this image and a resaved version compressed at different JPEG

qualities in the range [35,85]. At the originally saved quality of 65, the central region has a lower difference than the remaining image.

In comparison to [12], our approach does not require an estimate of

the discrete cosine transform (DCT) quantization from an assumed

original part of the image. Estimating the quantization from only

the underlying DCT coefcients is computationally nontrivial, and

prone to some estimation error, which leads to vulnerabilities in

the forensic analysis. In comparison to [13], our approach does

not require that the image be cropped in order to detect blocking

inconsistencies. In addition, our approach can detect local tampering

unlike the global approach of [13] which can only detect an overall

crop and recompression. And in comparison to [14], our approach,

although likely not as powerful, is computationally much simpler

and does not require a large database of images to train a support

vector machine (SVM). As with all forensic analysis, each technique

has its relative benets and drawbacks. The new technique described

here contributes to the growing set forensic analysis tools based on

JPEG artifacts, and should prove useful as a new tool in the arsenal

of forensic analysts.

II. JPEG G HOSTS

In the standard JPEG compression scheme [16], [17], a color image

(RGB) is rst converted into luminance/chrominance space (YCbCr).

The two chrominance channels (CbCr) are typically subsampled by a

factor of two relative to the luminance channel (Y). Each channel is

then partitioned into 8 8 pixel blocks. These values are converted

from unsigned to signed integers (e.g., from [0, 255] to [ 128, 127]).

Each block is converted to frequency space by using a 2-D DCT. Each

DCT coefcient is then quantized by an amount

round (1)

where the quantizationdepends on the spatial frequency and channel.

Larger quantization values yield better compression at the cost of

image degradation. Quantization values are typically larger in the

chrominance channels, and in the higher spatial frequencies, roughly

modeling the sensitivity of the human visual system.

156

IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 4, NO. 1, MARCH 2009

Fig. 3. Representative examples are shown from the 1 000 UCID images.

Fig. 4. ROC curves are shown for (a) a tampered region of size 150 150 and a quality difference of 15 and (b) a tampered region of size 100 100 and a quality

difference of 10. The solid curve corresponds to the accuracy of detecting the tampered region, and the dashed curve corresponds to the accuracy of correctly

classifying an authentic image. The vertical dotted lines denote (from left to right) false positive rates of 10%, 5%, and 1%. See also Table I.

Consider now a set of coefcients quantized by an amount ,

which are subsequently quantized a second time by an amount to

yield coefcients . With the exception of (i.e., no quantiza-

tion), the difference between and will be minimal when .

It is obvious that the difference between and increases for a quan-

tization value since the coefcients become increasingly more

sparse as increases. For values of , the difference between

and also increases because although the second quantization does

not affect the granularity of the coefcients, it does cause a shift in their

values. Shown in Fig. 1(a), for example, is the sum of squared differ-

ences between and as a function of the second quantization ,

where and where the coefcients are drawn from a normal

zero-mean distribution. Note that this difference increases as a function

of increasing , with the exception of , where the difference

is minimal. If is not prime, as in our example, then multiple minima

may appear at quality values that are integer multiples of . As will

be seen, this issue can be circumvented by averaging over all of the

JPEG DCT coefcients.

Consider now a set of coefcients quantized by an amount ,

followed by quantization by an amount to yield . Further

quantizing by yields the coefcients . As before, the difference

between and will be minimal when . But since the coef-

cients were initially quantized by , where , we expect to nd

a second minimum when . Fig. 1(b) shows the sum of squared

differences between and , as a function of , where and

. As before, this difference increases as a function of increasing

reaches a minimum at , and most interestingly, has a

second local minimum at . We refer to this second min-

imum as a JPEG ghost since it reveals that the coefcients were previ-

ously quantized (compressed) with larger quantization (lower quality).

Recall that the JPEG compression scheme separately quantizes each

spatial frequency within an 8 8 pixel block. One approach to detect

JPEG ghosts would be to separately consider each spatial frequency in

each of the three luminance/color channels. However, recall that mul-

tiple minima are possible when comparing integer multiple quantiza-

tion values. If, on the other hand, we consider the cumulative effect

of quantization on the underlying pixel values, then this issue is far

less likely to arise (unless all 192 quantization values at different JPEG

qualities are integer multiples of one another—an unlikely scenario 1 ).

1 The MPEG video standard typically employs JPEG quantization tables that

are scaled multiples of one another. These tables may confound the detection of

JPEG ghosts in MPEG video.

IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 4, NO. 1, MARCH 2009

157

Fig. 5. Original (left) and doctored (right) images are shown. The difference images at qualities 60 through 98 in steps of 2 are shown.

Therefore, instead of computing the difference between the quantized

DCT coefcients, we consider the difference computed directly from

the pixel values, as follows:

TABLE I

JPEG G HOST D ETECTION A CCURACY (%)

(2)

where denotes each of three RGB color chan-

nels, 2 and is the result of compressing at quality.

Shown in the top left panel of Fig. 2 is an image whose central

200 200 pixel region was extracted, compressed at a JPEG quality

of 65/100, and reinserted into the image whose original quality was 85.

Shown in each subsequent panel is the sum of squared differences (2)

between this manipulated image, and a resaved version compressed at

different JPEG qualities. Note that the central region is clearly visible

when the image is resaved at the quality of the tampered region (65).

Also note that the overall error reaches a minimum at the saved quality

of 85. There are some variations in the difference images within and

outside the tampered region which could possibly confound a forensic

analysis. These uctuations are due to the underlying image content.

Specically, since the image difference is computed across all spatial

frequencies, a region with small amounts of high spatial frequency con-

tent (e.g., a mostly uniform sky) will have a lower difference compared

to a highly textured region (e.g., grass). In order to compensate for these

differences, we consider a spatially averaged and normalized difference

measure. The difference image is rst averaged across a pixel re-

gion

(3)

and then normalized so that the averaged difference at each location

is scaled into the range [0,1]

2 The detection of JPEG ghosts is easily adapted to grayscale images by simply

computing (2) over a single image channel.

(4)

158

IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 4, NO. 1, MARCH 2009

Fig. 6. Original (left) and doctored (right) images are shown. Difference images at qualities 60 through 100 in steps of 2 are shown below.

Although the JPEG ghosts are often visually highly salient, it

is still useful to quantify whether a specied region is statistically

distinct from the rest of the image. To this end, the two-sample

Kolmogorov–Smirnov (K–S) statistic [18] is employed to determine

whether the distribution of differences (4) in two regions is similar or

distinct. The K–S statistic is dened as

a signicant impact. The second practical issue is that in the examples

shown before, we have assumed that the tampered region remains on

its original 8 8 JPEG lattice after being inserted and saved. If this

is not the case, then the misalignment may destroy the JPEG ghost

since new spatial frequencies will be introduced by saving on a new

JPEG block lattice. This problem can be alleviated by sampling all 64

possible alignments (a 0 to 7 pixel shift in the horizontal and vertical

directions). Specically, an image is shifted to each of these 64 loca-

tions prior to saving at each JPEG quality. Although this increases the

complexity of the analysis, each comparison is efcient, leading to a

minimal impact in overall run-time complexity.

(5)

where and are the cumulative probability distributions

of two specied regions in the computed difference , where

each value ofis considered separately.

There are two potential complicating factors that arise when de-

tecting JPEG ghosts in a general forensic setting. First, it is likely that

different cameras and photo-editing software packages will employ dif-

ferent JPEG quality scales and, hence, quantization tables [19]. When

iterating through different qualities, it would be ideal to match these

qualities and tables, but this may not always be possible. Working to

our advantage, however, is that the difference images are computed by

averaging across all spatial frequencies. As a result, small differences

in the original and subsequent quantization tables will not likely have

III. R ESULTS

To test the efcacy of detecting JPEG ghosts, 1 000 uncompressed

TIFF images were obtained from the Uncompressed Color Image Data-

base (UCID) [20]. These color images are each of size 512 384 and

span a wide range of indoor and outdoor scenes (Fig. 3). A central por-

tion from each image was removed, saved at a specied JPEG quality of

, reinserted into the image, and then the entire image was saved at the

same or different JPEG quality of . The MatLab function imwrite

was used to save images in the JPEG format. This function allows for

Plik z chomika:

Inne pliki z tego folderu:

Inne foldery tego chomika: