exposing doctored images by detecting duplicated and ... · matching image sift features to find...

9
Exposing Doctored Images by Detecting Duplicated and Distorted Regions Xunyu Pan and Siwei Lyu TR-CS-09-02 Computer Science Department College of Computing and Information University at Albany, SUNY Albany, NY 12222, USA {xypan,lsw}@cs.albany.edu Abstract We describe an efficient technique that automatically detects duplicated regions in a digital image for forensic analysis of its authenticity. Previous region duplication detection methods match pixel blocks to identify exact copies, and are not effective when the duplicated region undergoes geometric and illumination adjustments before being copied to the target location. Our method, on the other hand, can reliably detect such general cases of region duplication tampering. It works by matching image SIFT features to find potential duplicated regions, and is robust to geometric and illumination distortions of the duplicated regions. We quantify the performance of this technique with experimental evaluations, and show the efficacy on several credible forgeries.

Upload: dinhminh

Post on 26-Feb-2019

233 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Exposing Doctored Images by Detecting Duplicated and ... · matching image SIFT features to find potential duplicated regions, and is robust to geometric and illumination distortions

Exposing Doctored Images by Detecting Duplicated and Distorted Regions

Xunyu Pan and Siwei Lyu

TR-CS-09-02

Computer Science DepartmentCollege of Computing and Information

University at Albany, SUNYAlbany, NY 12222, USAxypan,[email protected]

Abstract

We describe an efficient technique that automatically detects duplicated regions in a digital image for forensic analysisof its authenticity. Previous region duplication detection methods match pixel blocks to identify exact copies, and are noteffective when the duplicated region undergoes geometric and illumination adjustments before being copied to the targetlocation. Our method, on the other hand, can reliably detect such general cases of region duplication tampering. It works bymatching image SIFT features to find potential duplicated regions, and is robust to geometric and illumination distortions ofthe duplicated regions. We quantify the performance of this technique with experimental evaluations, and show the efficacyon several credible forgeries.

Page 2: Exposing Doctored Images by Detecting Duplicated and ... · matching image SIFT features to find potential duplicated regions, and is robust to geometric and illumination distortions

1. IntroductionThanks to the increasing availability and sophistication

of digital imaging technology (digital cameras, computersand photo-editing software) and the Internet, digital photo-graphic images have become our main information source.However, concomitant with the ubiquity of digital imagesis the rampant problem of digital image tampering, whichhas seriously debased the credibility of photographic im-ages as definite records of events. One recent example isthe withdrawal of several photographs from the New YorkTimes magazine, which, in a statement issued by the Editor,“were digitally altered [...] and did not wholly reflect thereality they purported to show”1. To combat this problem,digital image forensics has emerged as a new research fieldthat aims to reveal digital tampering in images [9].

A common manipulation in tampering with an image,known as region duplication, is to copy and paste a con-tinuous portion of pixels to a different location in the sameimage. An example of tampered image with region duplica-tion is shown in Fig.1. The duplicated region is usually usedto alter the location of a person or an object, or to concealsome undesirable contents. Though conceptually very sim-ple, detecting duplicated region can be a challenging taskfor humans. Especially, experiments in psychophysics haveshown that searching identical regions (potentially under-gone geometric and illumination distortions) in an imagewith a lot of clutters and distractions is very hard for humansubjects [5]. Further, if the splicing of the duplicated regionis imperceptible, little concern is typically given to the factthat virtually identical regions are present in the image.

In recent years, quite a few techniques have been de-veloped to detect duplicated regions in images [11, 24, 22,15, 19], all of which try to find exact copies of small pixelblocks in an image. As such, they are most effective whenthe source region is copied and pasted to the target loca-tion with no change. However, in practice, such simplecopy-move alone can seldom create plausible tampered im-ages. More likely, as in the example of Fig.1, the dupli-cated regions are adjusted with geometric and illuminationtransforms (e.g., rotation, scaling and masking/boundarysmoothing) to better blend into the surroundings at the tar-get location. As such distortions change the spatial loca-tions and intensities of the pixels within the duplicated re-gions, methods based on direct matching of pixel blockswill not be very effective for their detection.

In this work, we describe a new technique that auto-matically detect duplicated regions in a digital image. Ourmethod collects SIFT keypoints as identifiers for duplicatedregions [21], employs robust matching of SIFT features toestimate transforms applied to the duplicated regions, and

1Reference source: http://www.nytimes.com/slideshow/2009/07/05/magazine/20090705-gilded-slideshow_index.html.

Figure 1: Top: An original image (courtesy of A. Popescu and H.Farid). Bottom: a forgery created with region duplication.

uses the estimated transform to identify the locations andextents of the source and the duplicated regions. Comparedto existing works, our method is a significant advancementproviding much more flexibility in detecting general regionduplication tampering. The main advantages of our methodcan be summarized as:

1. it can reliably detect duplicated regions that are geo-metrically distorted;

2. it is robust to general image degradations caused byJPEG compression or additive noise;

3. the algorithm has efficient running time – analyzing720× 436 images takes about 10 seconds.

We quantify the performance of this technique with experi-mental evaluations, and demonstrate its efficacy on severalcredible forgeries from various sources.

2. Related WorkThe overall goal of region duplication detection is to

identify all pixels in suspiciously similar regions in an im-age2. Several general techniques developed in digital imageforensics may be used to solve this problem. For instance,if both the original and the tampered image are in JPEG for-mat, many tampering operations (including region duplica-

2Though repeated patterns in an image (e.g., textures) can lead to falsepositives in duplication, it is assumed that in the original image there areno exact copies of such patterns.

Page 3: Exposing Doctored Images by Detecting Duplicated and ... · matching image SIFT features to find potential duplicated regions, and is robust to geometric and illumination distortions

tion) introduce double JPEG compression, whose effect canbe detected with specific statistical patterns of the quantizedDCT coefficients [25, 14]. However, slight perturbation tothe DCT coefficients of the tampered image can easily de-ceive this detection. One can also use more intrinsic cuesfrom the capture and production process of digital imagesfor tampering detection, such as the specific statistical pat-terns of interpolation [26] or color filter array demosaicing[27], and inconsistencies in camera response functions [20],lighting [17] or sensor noise [6, 29]. However, as the dupli-cated regions originate from the same image, their deviationfrom the properties of other image regions can be minimal,making these methods much less effective.

This motivates the development of detection methods de-signed specifically for region duplication. To date, most ex-isting detection methods, e.g., [11, 24, 22, 15, 19], focuson the simplest form of region duplication known as copy-move, where the source region is copied and pasted to thetarget location without any change. Detecting region copy-move can be achieved with a straightforward comparison ofpixel blocks to find exact copies. As a brute-force match ofall pixel blocks of a given size in an image has a runningtime quadratic to the size of the image, most methods focuson using low dimensional representations of pixel blocks,e.g., PCA [24, 22] or DCT [11, 19], and lexicographicalsorting to improve running efficiency. However, even withsuch accelerations, these block matching based methods canstill be slow in practice. More importantly, these methodsare not effective for more general duplication operationswhere the duplicated regions are subject to geometric andillumination adjustments.

Several recent works attempt to address this issue forsome special cases. In [15], invariant local image featuresare used to account for illumination changes to the dupli-cated regions in a copy-move tampering. In [18], the au-thors described a method that can detect copy-moved re-gions with boundary smoothing operations. The work in[19] can detect duplicated regions that have been rotated0, 90, 180 and 270. In our own previous work [4],we describe the first detection method that allows for moregeneral geometric transforms to the duplicated region, in-cluding rotation (with continuous angles) and scaling basedon matching image SIFT keypoints. The main drawback,however, is that the specific type of manipulation transform(rotation/scaling) has to be known prior to applying the al-gorithm.

3. MethodRegion duplication can be formalized as a 2D linear

transform between image regions. Denote pixel locationsin the source region and its duplication as ΩS and ΩT , re-spectively. Assuming only gentle changes in the pixel in-tensities in tampering the image, region duplication leads

ΩS

ΩT

T

image

Figure 2: A schematic illustration of region duplication operation.

to I(ΩT ) ≈ I(T (ΩS)), where T is a linear manipulationtransform including translation, rotation, scaling, perspec-tive, and their combinations (see Fig.2). Detecting regionduplication is to recover ΩS and ΩT , along with the ma-nipulation transform T . This problem bears resemblanceto the estimation of camera motion from multiple images[13]. The key differences are (1) here we are interested inthe transform between different regions in an image, and(2) we also need to recover the exact locations and extentsof the source and duplicated regions.

Our method detects duplicated regions in an image byfinding correspondences between keypoints, using them toestimate the manipulation transform, and locating all pixelsin source and duplicated regions using correlations com-puted with the estimated manipulation transform. In ourmethod, keypoint correspondences and the manipulationtransform are found by robust matching invariant image fea-tures using RANSAC. In what follows, we describe in detailhow keypoints are detected and putatively matched; howthese correspondences are used to compute the manipula-tion transform; and how this transform is used to output acorrelation map indicating pixels in duplicated regions.

3.1. Keypoint Detection and Putative Matching

The first step in our method is to identify keypoints,which are locations with distinct image information and ro-bust to scaling and rotation. We use the SIFT algorithm forthis purpose [21]. At each keypoint, a local image featurevector is generated from the normalized histograms of lo-cal gradients in a neighborhood of pixels of that keypoint.With the setting in [21], the SIFT features are 128 dimen-sional vectors at each keypoint.

We then seek putative matches for each SIFT keypoint.Specifically, we use the Best-Bin-First (BBF) algorithm [3]that matches a keypoint x with feature f to x′ with featuref ′, if f ′ is the nearest neighbor to f measured with their l2distance, and ‖f ′′ − f‖2/‖f ′ − f‖2 < ε, where f ′′ is thesecond nearest neighbor of f with regards to l2 distance,with ε a preset threshold. Furthermore, as we seek matchedkeypoints in one image, we have to avoid matching onekeypoint to itself. In particular, the nearest neighbor of akeypoint is searched outside a surrounding 11 × 11 win-dow of that keypoint. The leftmost panel of Fig.3 shows the

Page 4: Exposing Doctored Images by Detecting Duplicated and ... · matching image SIFT features to find potential duplicated regions, and is robust to geometric and illumination distortions

Figure 3: Key steps to detect duplicated regions. Left: putative matchings of SIFT keypoints. Middle: correlation map using the estimatedmanipulation transform. Right: duplication map after applying thresholds and morphological operations to the correlation map.

matched keypoints in a tampered image.

3.2. Estimating Manipulation Transform

Given a set of matched keypoints xi and x′i in an im-age, we can estimate the linear manipulation transform T .To allow for general affine transforms and translation tobe treated under the framework of linear transforms, weuse homogeneous coordinates for the matched keypoints asxi = (xi, yi, 1)T and x′i = (x′i, y

′i, 1)T , and represent the

manipulation transform by a planar homography betweenmatched keypoints xi ↔ x′i. This problem is then solvedby computing a 3×3 homography matrix H for each i suchthat:

Hxi = x′i, (1)

which is equivalent to a linear equation

[0T −w′ixT

i y′ixTi

w′ixTi 0T −x′ixT

i

] h1

h2

h3

= 0, (2)

where hj is the transpose of the j-th row of matrix H.In practice, with n ≥ 4 keypoint correspondences xi ↔x′i, the homography matrix H can be obtained using leastsquares estimation [13]. However, as many of the putativematching of keypoints obtained with BBF are mismatches,robust estimation methods such as RANSAC [10] are usedtogether with least squares estimation to alleviate the effectsof outliers.

In practice, we observe that subsequent detection perfor-mance relies on a very accurate estimation of the translationshift vector, and a small misalignment will introduce signif-icant errors. This is especially the case when the manipula-tion transform is close to pure translation (copy-move). Toget a more accurate detection for this special case, once wehave the estimated manipulation transform, we check if itis close to a mere translation (i.e., the affine part of H isclose to an identity. If so, we compute the l2 distances be-tween the location of each pair of matched SIFT keypoints.Though corresponding keypoints in the source and dupli-cated regions should have the same l2 distance, due to mis-match there is a range. So we build a histogram of pairwise

l2 distances, and collecting keypoint pairs with the distanceof maximum frequency of occurrence. The shift vector isthen estimated as the difference between the means of thesetwo sets of keypoints.

3.3. Obtaining Duplication Map

With the estimated manipulation transform, the sourceand the duplicated regions can be recovered by correlatingthe tampered image with its correction using the manipula-tion transform. Note the algorithm cannot differentiate the“source’ from the “duplicate”, but shows that they are iden-tical. Identical regions lead to high correlation in such acomparison. In detail, we first segment image into overlap-ping blocks of 4×4 pixels. We then transform the tamperedimage with the estimated H, and compute the correlationcoefficient between each pair of the corresponding 4 × 4blocks. The result, which we term as the correlation map,is shown in the middle panel of Fig.3.

We then process the correlation map to obtain an esti-mated contour for the source and the duplicated regions.Specifically, we first apply a Gaussian filter of size 7× 7 tosmooth the correlation map. Next, with a pre-chosen thresh-old, we binarize the correlation map. This is followed by re-moval of regions with areas smaller than another pre-giventhreshold to eliminate small isolated regions. The contoursof the potential source and duplicated regions are then con-nected with mathematical morphological operations to theduplication regions that (1) dilate then eroded to eliminateholes in the detected regions, and (2) eroded then dilated tosmooth the region contours [28]. As the last step, the recov-ered source and duplicated regions are combined togetherand mapped back to the original image coordinates. Anexample of the duplication map is shown in the rightmostpanel in Fig.3.

For RGB color images, duplicated regions are detectedfor each color channel individually, and we report the unionof detected regions of all three color channels. If there aremultiple duplication regions in the image, our algorithmwill iterate the overall process of keypoints matching, ma-nipulation transform estimation and generating duplicationmap, with keypoints in the duplicated regions already found

Page 5: Exposing Doctored Images by Detecting Duplicated and ... · matching image SIFT features to find potential duplicated regions, and is robust to geometric and illumination distortions

masked from further match. The whole algorithm stops un-til there is no further duplicated regions found.

4. Experimental ResultsIn this section, we report experimental results that test

the efficacy of the proposed region duplication detectionmethod3. We first demonstrate the detection results of ourmethod on several credible forgeries from various sources.We then quantify the general performance and robustnessof our method on a set of automatically generated tamperedimages with duplicated and distorted region.

In Fig.4, we show five representing cases of tamperedimages using region duplication, along with their detec-tions using our method as filled regions. The five casescorrespond to (1) simple copy-move, (2) illumination ad-justment, (3) scaling of the duplicated region with factor1.2, (4) rotation of the duplicated region of 60, and (5)a free form affine transform (combination of rotation andscaling) of the duplicated region. We create these forgeriesmanually using Photoshop, and take particular care in se-lecting source regions and finessing the duplicated regionto achieve convincing splicing effects. To provide a quanti-tative measure of performance, we define the pixel detectionaccuracy (PDA) and pixel false positive (PFP): denote ΩS

and ΩT as the detected regions corresponding to the sourceand its duplication,ΩS and ΩT , respectively:

pPDA =|ΩS

⋂ΩS |+ |ΩT

⋂ΩT |

|ΩS

⋃ΩT |

(3)

and

pPFP =|ΩS − ΩS |+ |ΩT − ΩT |

|ΩS

⋃ΩT |

. (4)

PDA quantifies the fraction of pixels in ΩS and ΩT that arecorrectly detected, and PFP evaluates the fraction of untam-pered pixels detected as from duplicated regions.

In these examples, and those described in the following,we set parameters as following: the correlation similaritythreshold for contour blocks is 0.3, the area threshold is500 pixels, and the nearest neighbor threshold is 0.5. AsFig.4 shows, our method is effective in revealing the loca-tions and extents of the source and the duplicated regions.Such are also reflected quantitatively from the pixel detec-tion accuracies and pixel false positive rates. Note that thelast three cases (scaling, rotation and shearing) will be par-ticular challenging for previous methods that are based onmatching pixel blocks directly, and the last case with shear-ing cannot be properly handled with our previous work [4]that can only detect pure rotation or scaling distortion. Asa simple sanity check, running our detection algorithm on

3Image data and code reproducing these results will be available on-line.

Figure 7: Top: an image doubted for digital tampering (courtesyof New York Times). Bottom: detected duplicated regions usingour method. This figure is better viewed in color.

the original image does not show any false detection. Over-all, the algorithm is also very efficient: for these 720× 436images the average running time is about 10 seconds.

Even with the aid of sophisticated photo-editing software(e.g., Photoshop), creating a convincing forgery with regionduplication usually requires a painstaking and laboriousmanual process. To facilitate such operations, recent yearsthere have been active developments in Computer Graph-ics for creating seamless region splicing, e.g., [23, 16, 1, 8],and have been featured in photo-editing tools [12, 2]. Tofurther test the efficacy of our detection method, we applyit to high quality forgeries that are created automatically us-ing such methods. Show in Fig.5 and 6 are the original im-ages, the forgeries and detection results using our method.The forgeries in Fig.5 are created using a coordinate-basedapproach that optimally render the transition between du-plicated region and the surroundings at the target location,leading to highly convincing tampering results. For theseimages, our method can recover fairly accurately the du-plicated regions, and correctly accounts for the geometricaland illumination distortions (e.g., rotation and scaling, aswell as hue changes that are most visible in the deer exam-ple). Note that our detection does not spot duplicated regionthat is reflected (second left antler of the deer) because SIFTfeatures are not invariant to such transform. But this may becircumvented if we include a reflected version of the tam-pered image. The tampered image in Fig.6 is created withthe Smart Fill tool developed by Alien Skin Software, whichrepresents the state-of-the-art in practical image retouchingtechniques. The tampered image has duplications of manysmall patches subtly adjusted and arranged to fit the contextat the target location, and does not have large and continu-ous duplicated regions. For this reason, it is extremely hardto find such duplications visually. On the other hand, usingour method, we are still able to single out some duplicatedregions with relatively large areas. However, the detectionis less reliable as we have smaller duplicated regions thatprovide less keypoints to match.

As another set of test, we run our detection method on animage that has appeared on the front pages of several inter-nationally important newspapers including The Los AngelesTimes, The Financial Times, and The Chicago Tribune andseveral major news Web sites early this year, Fig.7. Shortly

Page 6: Exposing Doctored Images by Detecting Duplicated and ... · matching image SIFT features to find potential duplicated regions, and is robust to geometric and illumination distortions

original image copy-move illumination

pPDA = 91.3%, pPFP = 0.3% pPDA = 83.8%, pPFP = 3.1%

scaling rotation shear

pPDA = 92.3%, pPFP = 0.8% pPDA = 88.0%, pPFP = 0.6% pPDA = 79.9%, pPFP = 1.9%

Figure 4: Detecting region duplication using our method. pPDA and pPFP stand for pixel detection accuracy and pixel false positive rate,respectively. See text for definition. This figure is better viewed in color.

after this image was published, doubts have been raised thatit has been digitally altered, a fact later confirmed by in-spection of photography experts and the appearance of an-other photography that was taken under similar condition4.We use our detection method to analyze this image, and areable to find two major region duplications in this image.Consistent with the expert analysis by visual inspection, ourmethod provides an objective and quantitative evaluation.

We further quantify the overall quality and robustnessof the detection based on homography estimation. For this

4http://thelede.blogs.nytimes.com/2008/07/10.

purpose, we prepare a database of 100 color images of size512 × 512 pixels. These images were cropped from larger2000× 3008 images taken with a Nikon D100 digital cam-era. In each image, a random square region of size 64× 64(corresponding to less than 1.6% of the image area) wascopied and pasted onto a random non-overlapping positionin the image, with proper clipping if the duplicated regiongoes beyond the image boundary. In creating these im-ages, we consider three manipulation transform (1) trans-lation: this corresponds to the simple copy-move; (2) rota-tion & translation: the rotating angle is randomly chosen in

Page 7: Exposing Doctored Images by Detecting Duplicated and ... · matching image SIFT features to find potential duplicated regions, and is robust to geometric and illumination distortions

Figure 5: Detecting region duplication in forgeries generated with algorithm in [8]. Left: original images. Middle: tempered images.Right: detection using the proposed method. The original and the tampered images are courtesy of the authors of [8]. This figure is betterviewed in color.

Figure 6: Detecting region duplication in forgeries generated using the Smart Fill tool of Alien Skin Software LLC [2]. Left: originalimages. Middle: tempered images. Right: detection using the proposed method. The original and the tampered images are courtesy ofAlien Skin Software LLC. This figure is better viewed in color.

the range of [0, 360), (3) scaling & translation: the scal-ing factor is randomly chosen in the range of [0.8, 2.0]5.As a comparison, we also apply our previous SIFT basedtransform-specific detection method [4] to these images.

Shown in Fig.8 are ROC curves of the three manipula-tion transforms, averaged over all test images. Note thatin all these cases and for both methods, for 20% false

5Scaling factors too large or too little are unlikely to be used in practicedue to the more visible interpolation artifacts.

PFP, the detection accuracies are around or above 80%, andachieve nearly perfect detection of the duplicated regions.We should point out that that both PDA and PDF are pixel-level measures, and we found that for reliable detection ofduplicated regions, performances of this level are usuallysufficient. In the case of rotation, the homography basedmethod outperforms the transform-specific detection, as thelatter requires an accurate estimate of the rotation angle,which is very hard with matched keypoints. On the other

Page 8: Exposing Doctored Images by Detecting Duplicated and ... · matching image SIFT features to find potential duplicated regions, and is robust to geometric and illumination distortions

simple copy-move scaling rotation

0 0.2 0.4 0.6 0.8 10

0.2

0.4

0.6

0.8

1

PFP

PD

A

proposed methodtransform−specific

0 0.2 0.4 0.6 0.8 10

0.2

0.4

0.6

0.8

1

PFP

PD

A

proposed methodtransform−specific

0 0.2 0.4 0.6 0.8 10

0.2

0.4

0.6

0.8

1

PFP

PD

A

proposed methodtransform−specific

Figure 8: ROC curves for different detection methods. See text for details

JPEG Q = 60 Q = 70 Q = 80 Q = 90 Q = 100mean 68.47/2.57 75.79/2.63 80.52/2.71 82.25/2.65 87.72/3.13std 20.01/1.97 17.78/1.60 12.06/1.77 12.82/1.58 7.54/1.90

SNR 20 dB 25 dB 30 dB 35 dB 40 dBmean 80.48/2.53 85.03/2.84 87.31/2.98 87.70/3.11 87.43/3.12std 14.24/1.34 9.49/1.57 5.70/1.69 5.62/1.98 6.10/1.74

Table 1: Means and standard deviations of PDA/PFP (in percent-age, over 100 images) for detecting duplication of 64×64 regionswith random rotations under different JPEG qualities and additivewhite Gaussian noise SNRs.

hand, the robust matching based on RANSAC used in thehomography-based method provides much reliable detec-tion. For scaling, both methods achieve comparable perfor-mance. It is for pure translation that the homography-baseddetection is not as effective, of which the reason is given inSection 3.2. As we have mentioned, if the affine part of theestimated manipulation transform is seen close to an iden-tity, we will run the copy-move detection method directlyfor optimal performance.

We further address the robustness and sensitivity of ouralgorithm by repeating the previous experiments with dif-ferent JPEG quality factors, or additive noise of differentsignal to noise ratios (SNR), for detecting of region du-plication with rotations. Shown in Table 1 are the meanand standard deviation of PDA/PFP as a function of JPEGcompression quality (top) and SNR of additive white Gaus-sian noise. The performance, in general, stays stable as theeffects of JPEG qualities and additive noise increase. Butalso, severe degradation (high noise or low JPEG compres-sion quality) can also render decrement in performance.

5. Conclusion and Future Work

Region duplication is a common tampering with digitalimages. Here, we describe a new technique to detect re-gion duplication tampering in digital images based on ro-bust matching of invariant image features. Compared toexisting methods, our method is effective to the detectionof duplicated regions undergone geometric and illumina-

tion adjustments. Experimental results with several credibleforgeries demonstrate the efficacy of our method.

There are several limitations of the proposed method.First, detection performance reduces as the size of the dupli-cated region gets smaller, for the reason that smaller regionsprovide less keypoints for matching (see Fig.6). Similarly,if the duplicated region has less visual structure (e.g., sky),the lack of reliable keypoints will also render our methodless effective. Second, prevalent repetitive structures or tex-tures may introduce false positives in detection. Third, cer-tain distortion of the duplicated region (e.g., extreme scalingor reflection) can break the invariance of the SIFT featuresand affect detection performance.

We are currently working on several extensions of thiswork. First, we only consider the duplicated region to be atransformed region in the original image. Recent develop-ment in Computer Graphics also suggests the possibility ofcreating duplicated regions from original image with statis-tical similarity (e.g., texture synthesis) [7]. It is thus of ourinterest to extend our method for the detection of such cases.Second, by incorporating temporal correlation, we are de-veloping similar detection method for region duplicationsin videos [30]. Finally, we are also in the process of con-verting our current algorithm into a plug-in for Photoshop,so that it can become part of the image forensic examiner’stoolkit.

References[1] A. Agarwala. Efficient gradient-domain compositing using

quadtrees. In ACM SIGGRAPH, 2007. 4[2] Alien Skin Software LLC, www.alienskin.com. Image Doc-

tor 2: Restore, Retouch, Remove and Repair. 4, 6[3] J. Beis and D. Lowe. Shape indexing using approximate

nearest-neighbour search in high-dimensional spaces. InCVPR, 1997. 2

[4] (blind review). 2, 4, 6[5] M. Bravo and H. Farid. Recognizing and segmenting objects

in clutter. Vision Research, 44(4):385–396, 2004. 1[6] M. Chen, J. Fridrich, M. Goljan, and J. Lukas. Determining

image origin and integrity using sensor noise. IEEE Trans-

Page 9: Exposing Doctored Images by Detecting Duplicated and ... · matching image SIFT features to find potential duplicated regions, and is robust to geometric and illumination distortions

actions on Information Forensics and Security, 3(1):74–90,March 2008. 2

[7] A. A. Efros and W. T. Freeman. Image quilting for texturesynthesis and transfer. Proceedings of SIGGRAPH 2001,pages 341–346, August 2001. 7

[8] Z. Farbman, G. Hoffer, Y. Lipman, D. Cohen-Or, andD. Lischinski. Coordinates for instant image cloning. ACMTrans. Graph., 28(3):1–9, 2009. 4, 6

[9] H. Farid. Photo fakery and forensics. In Advances in Com-puters. 2009. (to appear). 1

[10] M. A. Fischler and R. C. Bolles. Random sample consensus:a paradigm for model fitting with applications to image anal-ysis and automated cartography. Commun. ACM, 24(6):381–395, 1981. 3

[11] J. Fridrich, D. Soukal, and J. Lukas. Detection of copy-moveforgery in digital images. In Digital Forensic Research Work-shop, 2003. 1, 2

[12] T. Georgiev. Photoshop healing brush: a tool for seam-lesscloning. In Workshop on Applications of Computer Vission(ECCV), 2004. 4

[13] R. I. Hartley and A. Zisserman. Multiple View Geometry inComputer Vision. Cambridge University Press, 2nd edition,2004. 2, 3

[14] J. He, Z. Lin, L. Wang, and X. Tang. Detecting doctored jpegimages via dct coefficient analysis. In ECCV, 2006. 2

[15] H. Huang, W. Guo, and Y. Zhang. Detection of copy-moveforgery in digital images using SIFT algorithm. In IEEEPacific-Asia Workshop on Computational Intelligence andIndustrial Application, 2008. 1, 2

[16] J. Jia, J. Sun, C.-K. Tang, and H.-Y. Shum. Drag- and-droppasting. In ACM SIGGRAPH, 2006. 4

[17] M. Johnson and H. Farid. Exposing digital forgeries in com-plex lighting environments. IEEE Transactions on Informa-tion Forensics and Security, 3(2):450–461, 2007. 2

[18] D. Letscher. Detecting filtered cloning in digital images. InACM Workshop on MM&Sec, 2007. 2

[19] H.-J. Lin, C.-W. Wang, and Y.-T. Kao. Fast copy-moveforgery detection. WSEAS Trans. on Sig. Proc., 5(5):188–197, 2009. 1, 2

[20] Z. Lin, R. Wang, X. Tang, and H. Shum. Detecting doctoredimages using camera response normality and consistency. InCVPR, 2005. 2

[21] D. Lowe. Distinctive image features from scale-invariantkeypoints. IJCV, 60(2):91–110, 2004. 1, 2

[22] W. Luo, J. Huang, and G. Qiu. Robust detection of region-duplication forgery in digital images. In ICPR, 2006. 1, 2

[23] P. Perez, M. Gangnet, and A. Blake. Poisson image editing.In ACM SIGGRAPH, 2003. 4

[24] A. Popescu and H. Farid. Exposing digital forgeries by de-tecting duplicated image regions. Technical Report TR2004-515, Department of Computer Science, Dartmouth College,2004. 1, 2

[25] A. Popescu and H. Farid. Statistical tools for digital foren-sics. In 6th International Workshop on Information Hiding,Toronto, Canada, 2004. 2

[26] A. Popescu and H. Farid. Exposing digital forgeries by de-tecting traces of re-sampling. IEEE Transactions on SignalProcessing, 53(2):758–767, 2005. 2

[27] A. Popescu and H. Farid. Exposing digital forgeries in colorfilter array interpolated images. IEEE Transactions on SignalProcessing, 53(10):3948–3959, 2005. 2

[28] S. Suzuki and K. Abe. Topological structural analysis of dig-ital binary images by border following. CVGIP, 30(1):32–46, 1985. 3

[29] A. Swaminathan, M. Wu, and K. Liu. Digital image forensicsvia intrinsic fingerprints. IEEE Transactions on InformationForensics and Security, 3(1):101–117, March 2008. 2

[30] W. Wang and H. Farid. Exposing digital forgeries in videoby detecting duplication. In ACM Workshop on MM&Sec,2007. 7