IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 6, NO.

3, SEPTEMBER 2011 1099

A SIFT-Based Forensic Method for Copy–Move

Attack Detection and Transformation Recovery
Irene Amerini, Lamberto Ballan, Student Member, IEEE, Roberto Caldelli, Member, IEEE,
Alberto Del Bimbo, Member, IEEE, and Giuseppe Serra

Abstract—One of the principal problems in image forensics is

determining if a particular image is authentic or not. This can be
a crucial task when images are used as basic evidence to influence
judgment like, for example, in a court of law. To carry out such
forensic analysis, various technological instruments have been de-
veloped in the literature. In this paper, the problem of detecting if
an image has been forged is investigated; in particular, attention
has been paid to the case in which an area of an image is copied
and then pasted onto another zone to create a duplication or to
cancel something that was awkward. Generally, to adapt the image Fig. 1. Example of image tampering that appeared in press in July, 2008. The
patch to the new context a geometric transformation is needed. To feigned image (on the right) shows four Iranian missiles but only three of them
detect such modifications, a novel methodology based on scale in- are real; two different sections (encircled in red and purple, respectively) repli-
variant features transform (SIFT) is proposed. Such a method al- cate other image sections by applying a copy–move attack.
lows us to both understand if a copy–move attack has occurred
and, furthermore, to recover the geometric transformation used to
perform cloning. Extensive experimental results are presented to has been tampered with [6], [7] or which was the acquisition de-
confirm that the technique is able to precisely individuate the al- vice used [8], [9]. In particular, by focusing on the task of acqui-
tered area and, in addition, to estimate the geometric transforma- sition device identification, two main aspects must be studied:
tion parameters with high reliability. The method also deals with
multiple cloning. the first is to understand which kind of device generated a digital
image (e.g., a scanner, a digital camera, or a computer graphics
Index Terms—Authenticity verification, copy–move attack, dig- product) [10], [11], while the second is to determine which spe-
ital image forensics, geometric transformation recovery. cific camera or scanner (by recognizing model and brand) ac-
quired that specific content [8], [9].
The other main multimedia forensics topic is image tam-
I. INTRODUCTION pering detection [6], that is assessing the authenticity of a
digital image. Information integrity is fundamental in a trial,

D IGITAL crime, together with constantly emerging soft-

ware technologies, is growing at a rate that far surpasses
defensive measures. Sometimes a digital image or a video is
but it is clear that the advent of digital pictures and relative ease
of digital image processing today makes this authenticity uncer-
tain. Two examples of this problem, that recently appeared in
incontrovertible evidence of a crime or the proof of a malevo- newspapers and TV news, are given in Figs. 1 and 2. Modifying
lent action. By looking at digital content as a digital clue, multi- a digital image to change the meaning of what is represented
media forensics aims to introduce novel methodologies to sup- in it can be crucial when used in a court of law where images
port clue analysis and to provide an aid for making a decision are presented as basic evidence to influence the judgement.
about a crime. Multimedia forensics [1]–[3] deals with devel- Furthermore, it is interesting, once established that something
oping technological instruments operating in the absence of wa- has been manipulated, to understand exactly what happened: if
termarks [4], [5] or signatures inserted in the image. In fact, dif- an object or a person has been covered, if a part of the image
ferent from digital watermarking, forensics means are defined as has been cloned, if something has been copied from another
“passive” because they can formulate an assessment on a digital image, or if a combination of these processes has been carried
document by resorting only to the digital asset itself. These tech- out. In particular, when an attacker creates his feigned image
niques basically allow the user to determine if particular content by cloning an area of the image onto another zone (copy–move
attack), he is often obliged to apply a geometric transformation
to satisfactorily achieve his aim.
Manuscript received September 14, 2010; revised March 09, 2011; accepted
March 10, 2011. Date of publication March 17, 2011; date of current version In this paper, this issue is investigated, and the proposed
August 17, 2011. This work was supported in part by the EU ICT 3D-COFORM method is able to individuate if copy–move tampering has
Project (Contract FP7-231809). The associate editor coordinating the review of taken place and also to estimate the parameters of the trans-
this manuscript and approving it for publication was Dr. Wenjun Zeng.
The authors are with Media Integration and Communication Center, formation used (i.e., horizontal and vertical translation, scaling
University of Florence, 50134 Florence, Italy (e-mail: [email protected]; factors, rotation angle). On the basis of our preliminary work
[email protected]; [email protected]; [email protected]; serra@dsi. [12], a new methodology which satisfying these requirements
unifi.it).
Color versions of one or more of the figures in this paper are available online
is presented hereafter. Such a technique is based on the scale
at http://ieeexplore.ieee.org. invariant features transform (SIFT) [13], which are used to ro-
Digital Object Identifier 10.1109/TIFS.2011.2129512 bustly detect and describe clusters of points belonging to cloned

1556-6013/$26.00 © 2011 IEEE

1100 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 6, NO. 3, SEPTEMBER 2011

authors in [19] proposed a block representation calculated

using blur invariants. Their specific aim is to find features
invariant to the presence of blur artifacts that a falsifier can
apply to make detection of forgery more difficult. Then they
used principal component analysis (PCA) to reduce the number
of features and a -tree to identify the interesting regions. In
[20], the authors present a technique to detect cloning when
the copied part has been modified using two specific tools, the
Fig. 2. A close look at this picture, appearing in press in 2007 (Fars News
Agency, Tehran), shows that many elements are cloned over and over. Also in Adobe Photoshop healing brush and Poisson cloning. Another
this case the cloned sections are encircled in different colors. two algorithms, [16] and [21], are based on low dimensional
representations of blocks and fast sorting to improve efficiency
and have been developed to detect copy–move image regions.
areas. After detection, these points are exploited to reconstruct In particular, the authors in [16] apply a discrete cosine trans-
the parameters of the geometric transformation. The proposed form (DCT) to each block. Duplicated regions are then detected
technique has also been tested against splicing attacks (i.e., by lexicographically sorting the DCT block coefficients and
when an image block is duplicated onto a different image). In grouping similar blocks with the same spatial offset in the
fact, in a context where the source image is available (e.g., the image. In [21] the authors apply PCA on image blocks to yield
forensic analyst has to check a suspect dataset which contains a reduced-dimension representation. Duplicated regions are
both the source and the destination image), this methodology again detected by lexicographically sorting and grouping all
can be still applied. of the image blocks. A related approach is the method in [22]
The rest of the paper is structured as follows: Section II where a Fourier Mellin Transform is applied on each block.
presents related works regarding copy–move forgery detection Forgery decision is performed when there are more than a given
and it reviews the SIFT technique. Moreover, the contribution number of blocks that are connected to each other and the dis-
and the novelty of our approach respect to the state-of-the-art tance between block pairs is the same. To create a convincing
is discussed. Section III presents the proposed method in its forgery, it is often necessary to resize, rotate, or stretch portions
three main stages, while experimental results on forgery de- of an image. For example, when creating a composition of two
tection and on applied transformation parameters estimation objects, one object may have to be resized to match the relative
are presented in Section IV. Conclusions are finally drawn in heights. This process requires resampling of the original image
Section V. introducing specific periodic correlations between neighboring
pixels. The presence of these correlations due to resampling
II. SIFT FEATURES FOR IMAGE FORENSICS can be used to detect that something happened to the image
One of the most common image manipulations is to clone [23] but not to detect the specific manipulation.
(copy and paste) portions of the image, for instance, to conceal So a good copy–move forgery detection should be robust
a person or an object in the pictured scene. When this is done to some types of transformation, such as rotation and scaling,
with care, and retouching tools are used, it can be very difficult and also to some manipulations including JPEG compression,
to detect cloning. Moreover, since the copied parts are from the Gaussian noise addition, and gamma correction. Most existing
same images, some components (e.g., noise and color) will be methods do not deal with all these manipulations and are often
compatible with the rest of the image and thus will not be de- computationally prohibitive. In particular, the method in [21]
tectable using methods that look for incompatibilities in statis- is not able to detect scaling or rotation transformation, whereas
tical measures in different parts of the image [14], [15]. Further- with the methods in [16] and [22] only small variations in rota-
more, since the cloned regions can be of any shape and location, tion and scaling are identifiable as reported in [24]. The authors
it is computationally infeasible to search all possible image lo- in [25] make an attempt to overcome the problem using Zernike
cations and sizes with an exhaustive search as pointed out in moments to identify copy–move manipulation when only rota-
[16]. tion of the copied area takes place. This issue is also discussed
The problem of copy–move forgery detection has been faced in [26] where rotation transformations, JPEG compression, and
by proposing different approaches each of these based on the Gaussian noise manipulations are analyzed to understand how
same concept: a copy–move forgery introduces a correlation they affect the copy–move detection. The authors in [27] in-
between the original image area and the pasted one. Several stead propose a method to detect duplicated and transformed
methods search for this dependence by dividing the image regions through the use of a block description invariant to re-
into overlapping blocks and then applying a feature extraction flection and rotation such as the log-polar block representation
process in order to represent the image blocks with a low summed along its angle axis. Finally, a comparison among some
dimensional representation. In [17], the averages of red, green, of the copy–move methods described above has been reported
and blue components are chosen together with four other in [28], evaluating the performance of each method with and
features computed on overlapping blocks, and obtained by without geometric transformation applied to the copied patch.
calculating the energy distribution of luminance along four Today, local visual features (e.g., SIFT, SURF, GLOH, etc.)
different directions. A different approach is presented in [18] have been widely used for image retrieval and object recog-
in which the features are represented by the singular value nition, due to their robustness to several geometrical transfor-
decomposition (SVD) performed on low-frequency coefficients mations (such as rotation and scaling), occlusions, and clutter.
of the block-based discrete wavelet transform (DWT). The More recently, attempts have been made to apply these kinds of
AMERINI et al.: SIFT-BASED FORENSIC METHOD FOR COPY–MOVE ATTACK DETECTION AND TRANSFORMATION RECOVERY 1101

features also in the digital forensics domain; in fact, SIFT fea- the keypoint as the center and its canonical orientation as the
tures have been used for fingerprint detection [29], shoeprint origin axis. The contribution of each pixel is obtained by ac-
image retrieval [30], and also for copy–move detection [31], cumulating image gradient magnitude and orientation
[12], [32]. in scale-space and the histogram is computed as the
local statistics of gradient orientations (considering eight bins)
A. Review of the SIFT Algorithm in 4 4 subpatches.
Summarizing the above, given an image , this procedure
Most of the algorithms proposed in the literature for detecting ends with a list of keypoints each of which is completely
and describing local visual features usually require two steps. described by the following information:
The first is the detection step, in which interest points are local- where are the coordinates in the image plane, is the
ized, while in the second step robust local descriptors are built scale of the keypoint (related to the level of the image-pyramid
so as to be invariant with respect to orientation, scale, and affine used to compute the descriptor), is the canonical orientation
transformations. A comprehensive analysis of several local de- (used to achieve rotation invariance), and is the final SIFT
scriptors is provided in [33], while local affine region detectors descriptor.
are surveyed in [34]. These works confirm that SIFT features
[13] are a good solution because of their robust performance
and relatively low computational costs.
This method can be roughly summarized as the following B. Our Contribution
four steps: 1) scale-space extrema detection; 2) keypoint local-
ization; 3) assignment of one (or more) canonical orientations; A very preliminary work on copy–move forgery detection
4) generation of keypoint descriptors. based on SIFT features was proposed in [31], but in that paper
In other words, given an input image , SIFT features are no estimation of the parameters of the applied geometric trans-
detected at different scales using a scale-space representation formation is performed and, furthermore, extended numerical
implemented as an image pyramid. The pyramid levels are ob- results to evaluate real performances of the methodology (e.g.,
tained by Gaussian smoothing and subsampling of the image true/false positive rates) are not provided. Another very recent
resolution while interest points are selected as local extrema work has been presented in [32]. Although the technique is able
(minimum/maximum) in the scale-space. These keypoints, re- to deal with region extraction by resorting to a correlation map,
ferred to as in the following, are extracted by applying a com- it cannot manage affine transformation and, also in this case,
putable approximation of the Laplacian of Gaussian called Dif- quantitative results on the reliability of the estimate of geometric
ference of Gaussians (DoG). Specifically, a DoG image is transformation parameters are not given; in addition, this ap-
given by: proach adopts many different empirical thresholds whose set-
, where is the convolution ting seems to be not completely unsupervised. Moreover none
of the original image with the Gaussian blur of these contributions considers accurately the case of multiple
at scale . copy–move forgeries. As we will show furthermore, this is a key
In order to guarantee invariance to rotations, the algorithm point in a realistic forensic scenario since often a forged image
assigns to each keypoint a canonical orientation . To determine contains several cloned areas (like in the case of Fig. 2).
this orientation, a gradient orientation histogram is computed In this scenario is placed our proposed method that is able to
in the neighborhood of the keypoint. Specifically, for an image detect and then to estimate the geometrical transformation used
sample at scale (the scale in which that keypoint in a copy–move forgery attack. Multiple copy–move forgeries
was detected), the gradient magnitude and orientation are managed by performing a robust feature matching proce-
are precomputed using pixel differences dure and then a clustering on the keypoint coordinates in order
to separate the different cloned areas. These two tasks are fun-
damental since otherwise, in case of multiple cloning, it is often
impossible to detect and separate each forgery and also to es-
(1) timate the geometric transformation. Estimating the geometric
parameters with accuracy is deemed as a fundamental task not
(2)
only to understand how the cloned patch has been processed
[35] and possibly to infer which was the counterfeiter’s motive,
An orientation histogram with 36 bins is formed, with but also to compare the original source block of image and the
each bin covering approximately 10 . Each sample in the forged one on a common ground. Furthermore, a reliable esti-
neighboring window added to a histogram bin is weighted by mate of the transformation allows us to register the two patches
its gradient magnitude and by a Gaussian-weighted circular for a possible deeper forensics analysis [36]. The method pro-
window with equal to 1.5 times respect to the scale of the posed hereafter is able to deal with affine geometric transforma-
keypoint. The peaks in this histogram correspond to dominant tions and, as demonstrated by experimental results, also gives
orientations. Once these keypoints are detected, and canonical reliable estimates of the transformation parameters. Our tech-
orientations are assigned, SIFT descriptors are computed at nique works by relying on a unique empirical threshold which
their locations in both image plane and scale-space. Each regulates clustering operation, and that has been determined by
feature descriptor consists of a histogram of 128 elements, a training procedure on a general dataset. This is a very impor-
obtained from a 16 16 pixel area around the corresponding tant issue also in comparison with similar techniques like that
keypoint. This area is selected using the coordinates of in [32].
1102 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 6, NO. 3, SEPTEMBER 2011

Fig. 3. Overview of the proposed system. SIFT matched pairs and clusters.

III. PROPOSED METHOD very different from those of the rest of the set (i.e., features that
The proposed approach is based on the SIFT algorithm to ex- are globally distinctive). Therefore, the case of cloned patches
tract robust features which can allow it to discover if a part of is very critical since the keypoints detected in those regions are
an image was copy–moved, and furthermore, which geomet- very similar to each other.
rical transformation was applied. In fact, the copied part has For this reason, we propose a novel matching procedure that
basically the same appearance of the original one, thus key- is a generalization of (3), and is able to deal with multiple copies
points extracted in the forged region will be quite similar to of the same features. Our generalized 2NN test (referred as
the original ones. Therefore, matching among SIFT features can 2NN) starts from the observation that in a high dimensional
be adopted for the task of determining possible tampering. A feature space such as that of SIFT features, keypoints that are
simple schematization of the whole system is shown in Fig. 3: different from one considered share very high and very similar
the first step consists of SIFT feature extraction and keypoint values (in terms of Euclidean distances) among them. Instead,
matching, the second step is devoted to keypoint clustering and similar features show low Euclidean distances with respect to
forgery detection, while the third one estimates the occurred the others. The idea of the 2NN test is that the ratio between the
geometric transformation, if tampering has been detected. distance of the candidate match and the distance of the second
nearest neighbor is low in the case of a match (e.g., lower than
A. SIFT Features Extraction and Multiple Keypoint Matching 0.6) and very high in case of two “random features” (e.g., greater
Given a test image, a set of keypoints with than 0.6). Our generalization consists of iterating the 2NN test
their corresponding SIFT descriptors is extracted. between until this ratio is greater than (in our exper-
A matching operation is performed in the SIFT space among the iments we set this value to 0.5). If is the value in which the
vectors of each keypoint to identify similar local patches in procedure stops, each keypoint in correspondence to a distance
the test image. The best candidate match for each keypoint in (where ) is considered as a match for
is found by identifying its nearest neighbor from all the other the inspected keypoint.
keypoints of the image, which is the keypoint with the Finally, by iterating over keypoints in , we can obtain the set
minimum Euclidean distance in the SIFT space. In order to de- of matched points. All the matched keypoints are retained, but
cide that two keypoints match (i.e., “are these two descriptors isolated ones are no longer considered in subsequent processing
the same or not?”), simply evaluating the distance between two steps. Already at this stage a draft idea of the authenticity of
descriptors with respect to a global threshold does not perform the image is provided. But it can happen that images that legit-
well. This is due to the high-dimensionality of the feature space imately contain areas with very similar texture yield matched
(128) in which some descriptors are much more discriminative keypoints that might induce false alarms. The following two
than others. steps of the proposed methodology reduce this possibility.
We can obtain a more effective procedure, as suggested
in [13], by using the ratio between the distance of the closest B. Clustering and Forgery Detection
neighbor to that of the second-closest one, and comparing it with To identify possible cloned areas, an agglomerative hierar-
a threshold (often fixed to 0.6). For the sake of clarity, given a chical clustering [37] is performed on spatial locations (i.e.,
keypoint we define a similarity vector coordinates) of the matched points. Hierarchical clustering cre-
that represents the sorted Euclidean distances with respect ates a hierarchy of clusters which may be represented by a tree
to the other descriptors. Following this idea, the keypoint is structure. The algorithm starts by assigning each keypoint to
matched only if this constraint is satisfied: a cluster; then it computes all the reciprocal spatial distances
where (3) among clusters, finds the closest pair of clusters, and finally
merges them into a single cluster. Such computation is itera-
We refer to this procedure as the 2NN test. This matching pro- tively repeated until a final merging situation is achieved. The
cedure has one main drawback: it is unable to manage multiple way this final merging can be accomplished is basically condi-
keypoint matching. This is a key aspect in case of copy–move tioned both by the linkage method adopted and by the threshold
forgeries since it may happen that the same image area is cloned used to stop cluster grouping.
over and over (see, for example, Fig. 2). In other words, it only Several linkage methods exist in the literature and our ex-
finds matches between keypoints whose SIFT descriptors are periments evaluate their performance to estimate the best cutoff
AMERINI et al.: SIFT-BASED FORENSIC METHOD FOR COPY–MOVE ATTACK DETECTION AND TRANSFORMATION RECOVERY 1103

threshold for forgery detection (see Section IV-A for a de- a flat patch). Anyway, this is a very well-known open issue in
tailed description of such experiments). In particular, three dif- SIFT-related scientific literature.
ferent linkage methods have been evaluated: Single, Centroid,
and Ward’s linkage. Given two clusters and , respectively C. Geometric Transformation Estimation
containing and objects (where and indicate When an image has been classified as nonauthentic, the pro-
the th and the th object in the clusters and ), the linkage posed method can determine which geometrical transformation
methods operate as follows: was used between the original area and its copy–moved ver-
1) Single linkage uses the smallest Euclidean distance be- sion. Let the matched point coordinates be, for the two areas,
tween objects in the two clusters and , respectively; their geo-
metric relationships can be defined by an affine homography
which is represented by a 3 3 matrix as
(4)

2) Centroid linkage uses the Euclidean distance between the (9)

centroids of the two clusters

(5) This matrix can be computed by resorting to at least three

matched points. In particular, we determine using maximum
where likelihood estimation of the homography [38]. This method
seeks homography and pairs of perfectly matched points
and (6) and that minimizes the total error function as in (10)

3) Ward’s linkage evaluates the increment/decrement in the

error sum of squares (ESS) after merging the two clusters (10)
into a single one with respect to the case of two separated
clusters However, mismatched points (outliers) can severely disturb
the estimated homography. For this reason, we perform the pre-
(7)
vious estimation by applying the RANdom SAmple Consensus
where algorithm (RANSAC) [39]. This algorithm randomly selects a
set (in our case three pairs of points) from the matched points
(8) and estimates the homography ; then all the remaining points
are transformed according to and compared in terms of dis-
tance with respect to their corresponding matched ones. If this
where is the centroid (again) and indicates the distance is under or above a certain threshold , they are cat-
combined cluster. alogued as inliers or outliers, respectively. After a predefined
According to the adopted linkage method, a specific tree number of iterations, the estimated transformation which
structure is obtained. In addition to this, the proper choice is associated with the higher number of inliers is chosen. In our
of the threshold , to determine where to cut the tree and experimental tests, has been set to 1000 and the threshold
consequently which is the final number of clusters, is crucial. to 0.05; this is due to the fact that we used a standard method
The parameter which is compared with is the inconsistency of normalization of the data for homography estimation. The
coefficient (IC) which characterizes each clustering operation; points are translated so that their centroid is at the origin and
a higher value of this coefficient, the less similar the objects then they are scaled so that the average distance from the origin
connected by the link, thus when it exceeds the threshold is equal to . This transformation is applied to both of the two
clustering stops. IC takes basically into account the average areas and independently.
distance among clusters and does not allow us to join clusters Once the affine homography is found, rotation and scaling
spatially too far at that level of hierarchy. It is easy to see transformations can be computed from its decomposition, while
that an appropriate choice of directly influences tampering translation can be determined by considering the centroids of the
detection performance. At the end of the clustering procedure, two matched clusters. In particular, can be represented as
clusters which do not contain a significant number (more than
three) of matched keypoints are eliminated. On this basis, where (11)
to optimize detection performance and consequently to the
carried out experimental tests (see again Subsection IV-A), we
The matrix is the composition of rotation and nonisotropic
consider that an image has been altered by a copy–move attack
scaling transformations. In fact, it can always be decomposed as
if the method detects two (or more) clusters with at least three
pairs of matched points linking a cluster to another one. This (12)
aspect has been investigated and this assumption yields a good
trade-off between the need to provide a low false alarm rate. where and are rotations by and , respectively, and
It should be noted that it can happen that no matched key- is a diagonal matrix for the scaling trans-
points are obtained, mainly because salient features are not de- formation. Hence, the defines the concatenation of a rotation
tected in the forged patch (e.g., when an object is hidden with by , a scaling by and , respectively, in the rotated and
1104 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 6, NO. 3, SEPTEMBER 2011

TABLE I TABLE III

TEN DIFFERENT COMBINATIONS OF GEOMETRIC TRANSFORMATIONS APPLIED TRAINING PHASE: TPR AND FPR VALUES (IN PERCENTAGE) FOR EACH
TO THE ORIGINAL PATCH FOR THE MICC-F220 DATASET METRIC WITH RESPECT TO

TABLE II
FOURTEEN DIFFERENT COMBINATIONS OF GEOMETRIC TRANSFORMATIONS
APPLIED TO THE ORIGINAL PATCH FOR THE MICC-F2000 DATASET

applied to the and axis of the cloned image part (e.g., in the
attack , the and axes are scaled by 30%, and no rotation is
performed).

A. Settings for Forgery Detection

directions; a rotation back by ; and finally another rota- First of all, the proposed method is analyzed to determine
tion by . This decomposition is computed directly by the SVD. the best settings for the cutoff threshold introduced in
In fact, the matrix can be also rewritten as Section III-B according to the chosen linkage method. These
since and are orthog- values are set for the all remaining experiments and compar-
onal matrices. isons. To address this problem, the following experiment has
been setup applying a four-fold cross-validation process: from
IV. EXPERIMENTAL RESULTS the database of 220 images (MICC-F220), 165, that is of the
In this section, we evaluate the proposed methodology image set, (82 tampered and 83 original) have been randomly
providing two main kinds of the tests: first, on a small dataset chosen to perform a training to find the best threshold for
named MICC-F220, a benchmarking of the technique is done each of the three considered linkage methods (Single, Centroid,
to properly set the threshold and to compare it with other Ward’s); the remaining 55 images ( of the whole set) have
methods known in the literature; second, on a larger dataset been used in a successive testing phase to evaluate detection
named MICC-F2000, a complete evaluation is carried out performances of the proposed technique. The experiment was
by testing the system against different types of modifica- repeated four times, by cyclically exchanging the four image
tions. Both datasets are composed of images with different subsets belonging to the training (3 subsets) and to the testing
contents coming from the Columbia photographic image repos- set (1 subset), and the results have been averaged. Detection
itory [40] and from a personal collection. The first dataset performance was measured in terms of true positive rate (TPR)
MICC-F220 consists of 220 images: 110 are tampered im- and false positive rate (FPR), where TPR is the fraction of
ages and 110 are originals. The image resolution varies from tampered images correctly identified as such, while FPR is the
722 480 to 800 600 pixels and the size of the forged fraction of original images that are not correctly identified
patch covers, on the average, 1.2% of the whole image. The
second dataset MICC-F2000 is composed of 2000 photos of
2048 1536 pixels (3 M pixels) and the forgery is, on average,
1.12% of the whole image: so it is again quite small and similar
to the MICC-F220 dataset case. To reproduce as much as
possible a practical situation, the number of original and altered We repeat that an image is considered to have been altered by
images belonging to the MICC-F2000 dataset is not the same: a copy–move attack if the method detects two (or more) clusters
1300 original images and 700 tampered images have been with at least three pairs of matched points that link a cluster to
taken. The forged images are obtained, in both the datasets, another one (as detailed in Subsection III-B).
by randomly selecting (both as location and as dimension) an In Table III, for each linkage method, the TPR and the FPR
image area (square or rectangular) and copy-pasting it over obtained during the training phase are reported with respect to
the image after having applied a number of different attacks the threshold , which varies in the interval [0.8, 3] with steps
such as translation, rotation, scale (symmetric/asymmetric), or of 0.2. The goal was to minimize the FPR while maintaining a
a combination of them. very high TPR; we see that FPR is almost always very low, while
Tables I and II summarize the geometric transformations for on the contrary TPR is highly variable. The optimal threshold
the attack applied in the MICC-F220 dataset (10 attacks, from has been chosen, as evidenced in Table III, for the maximum
to in Table I) and in the MICC-F2000 (14 attacks, from value of TPR that is 1.6 for the Single linkage method, 1.8 for the
to in Table II), respectively. In particular, for each attack, we Centroid and 2.2 for the Ward’s linkage. Finally, the test phase
report the rotation in degrees and the scaling factors , has been launched for the best metrics using the previously
AMERINI et al.: SIFT-BASED FORENSIC METHOD FOR COPY–MOVE ATTACK DETECTION AND TRANSFORMATION RECOVERY 1105

TABLE IV 2) Copy–Move Methods Comparison: The proposed ap-

TEST PHASE ON MICC-F220 DATASET: DETECTION RESULTS proach has been compared to the results obtained with our
IN TERMS OF FPR AND TPR
implementations of the methods presented in [16], based on
DCT, and in [21], based on PCA (both have been previously
introduced in Section II). The input parameters required by the
two methods are set as follows: (number of pixels per
block), (number of neighborhood rows to search in
TABLE V the lexicographically sorted matrix), (threshold
TRANSFORMATION PARAMETER ESTIMATION ERRORS FOR THE MICC-F220 for the minimum frequency), and (threshold to
(SINGLE LINKAGE METHOD WITH , AS PREVIOUSLY UNDERLINED determine a duplicated block). These parameters are used in
OTHER METRICS GIVE SIMILAR PERFORMANCES). THE VALUES AND
ARE EXPRESSED IN PIXELS WHILE IS IN DEGREES
both the algorithms, while (fraction of the ignored
variance along the principle axes after PCA is computed) and
(number of the quantization bins) are only used for
the method presented in [21]. In our method, Ward’s linkage
with a threshold has been used.
The experiments have been launched on the whole
obtained in the training phase. The final detection results are re- MICC-F220 image database on a machine with an Intel
ported in Table IV. These values show that the proposed method Q6600 quad core with 4-GB RAM and the FPR, TPR, and
performs satisfactorily, providing a low FPR while maintaining processing time have been evaluated. Table VII shows the
a high rate of correct tampering detection for all the used linkage detection performance and the processing time on average (in
methods, though Ward’s metric seems to be slightly better. We seconds) for an image.
conclude that the choice of linkage methods is not so funda- The results indicate that the proposed method performs better
mental, while the setting is crucial. with respect to the others methods; in fact the processing time
Furthermore, for the cases of correctly detected forged im- (per image) is on average about 5 s, whereas the other two take
ages, the estimation of the geometric transformation parameters more than 1 min and almost 5 min, respectively. Furthermore,
which bring the original patch onto the forged one has also been the DCT and PCA methods, though obtaining an acceptable
computed. The mean absolute error (MAE) between each of the TPR, fail when a decision about original image is required (high
true values of the transformation parameters and the estimated FPR values in Table VII). This is basically due to the incapacity
ones are reported in Table V. As in the previous tables, and of such methods to properly deal with cases where a geomet-
refer to the scaling factors occurred in the transformation, rical transformation which is not pure translation is applied to
refers to the rotation (in degrees), while and are translation the copy–moved patch. For the specific case of pure patch trans-
on direction, respectively. lation, FPR is 0% for all the three methods.
Results show a high degree of precision in the estimate of the
various parameters of the affine transformation. In addition to
B. Test on Multiple Copied Regions
this, Table VI reports for one of the test images belonging to the
MICC-F220, named Cars (see Fig. 4, first column), each trans- In this experiment, we analyze the performance of our method
formation parameter (the original value applied to the patch, the in the presence of tampered images which have multiple copies
estimated one and the absolute error ). It can be observed of the same region. This test has been performed on ten photos
how reliable the estimate is, specifically for the scale parameters of 2048 1536 pixels. In these pictures, one or more image
and also for an asymmetric scaling combined with a rotation. areas are copied and pasted in several different positions over
1) Qualitative Evaluation: Here we report some experi- the image, taking into account the context in order to hide, at
mental results on images where a copy–move attack has been first glance, the forgery.
performed by taking into account the context. In this case, the In this scenario, as we have previously highlighted in
patch is selected according to the specific goal to be achieved Subsection III-A, the standard 2NN matching procedure is a
and, above all, transformed by paying attention to perfectly critical point for copy–move forgery detection methods based
conceal the modification. Alterations are not recognizable at on SIFT features [31], [32]. In fact, comparing the standard
least at a first glance and a forensic tool could help in investi- SIFT matching technique with our 2NN strategy, we see that
gations. In Fig. 4, four such cases are pictured by presenting our method increases (on average) by 195% the number of the
the tampered image and the corresponding one where matched extracted matches. A high number of matches is fundamental in
keypoints and clusters, extracted by the proposed method, are order to have sufficient information for a correct estimation of
highlighted. An interesting situation concerns the individuation the geometric transformation, but it can introduce false alarms.
of a cloned patch for the image named Dune (second column) To this end, we tested our matching strategy on the MICC-F220
where, though the duplicated area is quite flat, the method is dataset in order to evaluate how these new matches influence
able to detect a sufficient number of matched keypoints. On the the results. We found that we lose on average 3% in terms of
contrary, the opposite occurs for the image named Santorini FPR with the same results in terms of TPR.
(last column), where a large number of matched keypoints is Fig. 5 shows a qualitative comparison between the two tech-
obtained; now the cloned block is very textured and though niques. It is interesting to note that the number of the matched
it has undergone a geometrical transformation to be properly keypoints between the gun and , obtained by the standard
adapted to the context, the SIFT algorithm is robust enough not 2NN technique, are very few (2 matches) with respect to 2NN
to be disturbed. (54 matches). In this case, the technique, based on standard
1106 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 6, NO. 3, SEPTEMBER 2011

TABLE VI
TRANSFORMATION PARAMETER ESTIMATION ON IMAGE CARS. THE VALUES AND ARE EXPRESSED IN PIXELS WHILE IS IN DEGREES

Fig. 4. Some examples of tampered images are pictured in the first row; the corresponding detection results are reported in the second row.

TABLE VII method against usual operations such as JPEG compression or

TPR, FPR VALUES (%) AND PROCESSING TIME (AVERAGE, PER IMAGE) noise addition, an image can undergo, have been carried out;
FOR EACH METHOD
such kinds of processing have been considered as applied both
to the whole forged image and only to the altered image patch.
First of all, we set up again an experiment to determine the
best threshold , according to the three linkage methods, as
done in Section IV-A for the MICC-F220; this has been made to
further check if the established thresholds were correct. To do
this, a four-fold cross-validation process has been carried out.
Results are listed in Table VIII. It can be observed that a sim-
ilar behavior to that obtained with MICC-F220 is obtained and,
above all, that the values chosen in Subsection IV-A for (1.6
for Single, 1.8 for Centroid, and 2.2 for Ward’s) still yield higher
performance in terms of TPR and FPR. After this, the test phase
is launched by setting such values for . In Table IX, the de-
tection rates are reported demonstrating both the effectiveness
Fig. 5. Matched keypoints computed by the 2NN standard SIFT matching tech-
nique (left) and our 2NN strategy (right). of the proposed method which achieves a TPR of around 93%
for all three metrics, and its robustness obtaining again perfor-
mances very consistent with those in Table VIII for these fixed
thresholds.
matching, fails to detect the relationship between the two guns.
Going into detail, in Fig. 7, the number of errors for each
Finally, Fig. 6 shows some examples of multiple cloning ob-
attack is listed with regard to tampered images not detected as
tained with the 2NN test. Detection results are reported by
such. The most critical attacks seem to be the (
highlighting matched keypoints and clusters.
and ) and the ( and ) which
increase twice the patch dimension and apply a 40 rotation
C. Test on a Large Dataset
combined with a consistent variation in scale. The histogram
In this section, experimental results obtained on a larger in Fig. 7 shows that these two kinds of attacks generate around
dataset, named MICC-F2000, to verify the behavior of the 30% of the total errors.
proposed technique are presented; detection performances and In Table X, we report the errors in estimated geometric trans-
geometric transformation parameters estimation are investi- formation parameters averaged over all 500 test images. The
gated as well. Furthermore, tests to check the robustness of the MAE still remains small although the transformations applied
AMERINI et al.: SIFT-BASED FORENSIC METHOD FOR COPY–MOVE ATTACK DETECTION AND TRANSFORMATION RECOVERY 1107

Fig. 6. Examples of tampered images with multiple cloning are shown in the first row, while the detection results are reported in the second row.

Fig. 7. Error analysis of tampered images misdetection for each different attack (in percentage).

TABLE VIII TABLE X

TRAINING PHASE ON MICC-F2000 DATASET: TPR AND FPR VALUES FOR TRANSFORMATION PARAMETER ESTIMATION ERRORS. THE VALUES AND
EACH METRIC WITH RESPECT TO ARE EXPRESSED IN PIXELS, WHILE IS IN DEGREES

pact of JPEG compression and noise addition on all the 2000

images of the MICC-F2000 dataset has been investigated. In
the first experiment all the images which were originally in the
JPEG format (quality factor of 100), have been compressed in
JPEG format with a decreasing quality factor of 75, 50, 40, and
20. Table XI (left) shows the FPR and TPR (Ward’s linkage
method with ) for all the diverse JPEG quality fac-
tors; it can be seen that FPR is practically stable while the TPR
TABLE IX tends to slightly diminish when image quality decreases. In the
TEST PHASE ON MICC-F2000 DATASET: DETECTION RESULTS IN TERMS second experiment, in the same way as before, the images of
OF FPR AND TPR OBTAINED WITH , , AND MICC-F2000 dataset are distorted by adding a Gaussian noise
FOR THE THREE LINKAGE METHODS, RESPECTIVELY
to obtain a decreasing signal-to-noise ratio (SNR) of 50, 40, 30,
and 20 dB. Noisy images are obtained by adding white Gaussian
noise to the image with a JPEG quality factor of 100. In Table XI
(right), obtained results are shown and it can be noticed that the
TPR is over 90% until an SNR of 30 dB while FPR is again
to the images for the MICC-F2000 dataset are more challenging quite stable, though it seems to even improve.
with respect to those in the MICC-F220 dataset. 2) JPEG Compression, Noise Addition, and Gamma Correc-
1) JPEG Compression and Noise Addition: The proposed tion on Copied Patch: Duplicated patches are often modified
methodology has also been tested in terms of detection perfor- by applying further processing such as brightness/contrast ad-
mance from a robustness point of view; in particular, the im- justment, gamma correction, noise addition, and so on, in order
1108 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 6, NO. 3, SEPTEMBER 2011

TABLE XI TABLE XIII

DETECTION PERFORMANCES AGAINST JPEG COMPRESSION (LEFT) AND NOISE DETECTION PERFORMANCES AGAINST SPLICING ATTACK
ADDITION (RIGHT) (IN PERCENTAGE)

to create fraudulent content belongs to the image collection at

disposal. It is easy to understand that the proposed method can
be adopted again to determine both if within the to-be-checked
TABLE XII collection there is a false image containing an “external” patch
DETECTION PERFORMANCE AGAINST GAMMA CORRECTION, JPEG
COMPRESSION, AND NOISE ADDITION APPLIED TO THE DUPLICATED
and, above all, where it comes from. It is interesting to highlight
AND GEOMETRICALLY TRANSFORMED PATCH that succeeding in detecting such links could help investigation
activities. To show that the proposed technique can be used in
such a scenario the following experimental test has been set up.
A subset of 100 images (96 original and 4 tampered with)
taken from a private collection with size of 800 600 pixels has
been selected. In particular, the 4 fake images have been created
by pasting a patch that was cut from another image belonging to
to adjust the patch with respect to the image area where it has to the other original 96. The proposed technique has been launched
be located. To explore this scenario, the following experiment to analyze all the possible pairs of photos
has been performed. Starting from ten original images, a block in the dataset looking for duplicated
is randomly (as explained before) selected for each of them and areas. To allow to the presented algorithm to perform as is, the
four geometric transformations ( , , , and from Table II) pair of images to be checked are considered as a single image
are applied to every one of these patches. Furthermore, before with a double number of columns (size equal to );
pasting them, four different gamma corrections with values [2.2, due to this, the detection threshold has been increased up
1.4, 0.7, 0.45] are applied to each block. Finally, 160 tampered to 3.4 (it was 2.2 in the previous experiments of this section) for
images are obtained. In the same way, the final stage of gamma Ward’s linkage method which was chosen for this experiment.
correction is first substituted by JPEG compression with dif- In Table XIII, performances in terms of FPR and TPR are re-
ferent quality factors [75, 50, 40, 20] and second by Gaussian ported.
noise addition with SNR (dB) equal to 50, 40, 30, 20. For every The method is able to correctly reveal all four fake pairs, de-
case, 160 fake images have been created. So for each of the three termining a link between the possible original image and the
situations (gamma correction, JPEG compression, and noise ad- forged one, though it cannot distinguish the source from the
dition), a dataset composed by 160 fake images and by 350 orig- destination if other tools are not adopted. The procedure also
inal ones randomly taken from the MICC-F2000 database is detects two other innocent pairs of images causing false alarms.
built. In Table XII, performance in terms of TPR and FPR is In Fig. 8, the four cases of splicing attack detection are pictured,
reported. while in Fig. 9, one of the false alarms is illustrated. In this last
These experiments show that the proposed method maintains case, we see immediately that the error is caused by the presence
its level of accuracy though diverse kinds of postprocessing are of the same objects (the posters over the wall of the wooden box)
applied to the duplicated patch in addition to a geometric trans- in both the photos taken in the same real context. However, this
formation, to adapt it to the image context where it is pasted. could be an actual situation that happens in practical scenarios
(e.g., establishing possible relations among photos acquired in
D. Image Splicing similar environments).
Though the proposed technique has been presented to op-
V. CONCLUSION
erate in a copy–move attack scenario, it can also be utilized in a
context where a splicing operation has occurred. With the term A novel methodology to support image forensics investiga-
splicing attack we mean that a part of an image is grabbed and, tion based on SIFT features has been proposed. Given a sus-
possibly after having been adapted (geometrically transformed pected photo, it can reliably detect if a certain region has been
and/or enhanced), and pasted onto another one to build a new, duplicated and, furthermore, determine the geometric transfor-
fake, image. In most cases, only the final fake photo is avail- mation applied to perform such tampering. The presented tech-
able to the forensic analyst for inspection, the source image is nique shows effectiveness with respect to diverse operative sce-
often undeterminable; because of this, the SIFT matching pro- narios such as composite processing and multiple cloning. Fu-
cedure, which is the core of the proposed method cannot take ture work will be mainly dedicated to investigating how to im-
place and would seem that there is no room for it in such a cir- prove the detection phase with respect to the cloned image patch
cumstance. This is not always true in practice. In fact, often, the with highly uniform texture where salient keypoints are not re-
analyst is required to give an assessment of a dataset of images, covered by SIFT-like techniques. In particular, integration with
for example, belonging to a specific person under judgement, or other forensics techniques applied locally onto flat zones is en-
that have been found on a hard disk or a pen drive, and so on. In visaged. Furthermore, the clustering phase will be extended by
this operative scenario, it can happen that the source image used means of an image segmentation procedure.
AMERINI et al.: SIFT-BASED FORENSIC METHOD FOR COPY–MOVE ATTACK DETECTION AND TRANSFORMATION RECOVERY 1109

Fig. 8. Examples of correct detection of a splicing attack.

[13] D. G. Lowe, “Distinctive image features from scale-invariant key-

points,” Int. J. Comput. Vision, vol. 60, no. 2, pp. 91–110, 2004.
[14] S. Bayram, I. Avcibas, B. Sankur, and N. Memon, “Image manipula-
tion detection with binary similarity measures,” in Proc. EUSIPCO,
Antalya, Turkey, 2005.
[15] H. Farid and S. Lyu, “Higher-order wavelet statistics and their applica-
tion to digital forensics,” in Proc. IEEE CVPR Workshop on Statistical
Analysis in Computer Vision, Madison, WI, 2003.
[16] J. Fridrich, D. Soukal, and J. Lukás, “Detection of copy-move forgery
in digital images,” in Proc. DFRWS, Cleveland, OH, 2003.
Fig. 9. Example of wrong detection of a splicing attack. [17] W. Luo, J. Huang, and G. Qiu, “Robust detection of region-duplication
forgery in digital image,” in Proc. ICPR, Washington, DC, 2006.
[18] G. Li, Q. Wu, D. Tu, and S. J. Sun, “A sorted neighborhood approach
for detecting duplicated regions in image forgeries based on DWT and
SVD,” in Proc. IEEE ICME, Beijing, China, 2007.
ACKNOWLEDGMENT [19] B. Mahdian and S. Saic, “Detection of copy-move forgery using a
method based on blur moment invariants,” Forensic Sci. Int., vol. 171,
The authors would like to thank L. Del Tongo for his support no. 2–3, pp. 180–189, 2007.
in the preparation of the experiments. [20] B. Dybala, B. Jennings, and D. Letscher, “Detecting filtered cloning in
digital images,” in Proc. ACM Int. Workshop on Multimedia & Security
(MM&Sec), New York, 2007.
REFERENCES [21] A. Popescu and H. Farid, Exposing Digital Forgeries by Detecting Du-
plicated Image Regions Computer Science Dept., Dartmouth College,
[1] S. Lyu and H. Farid, “How realistic is photorealistic?,” IEEE Trans. Hanover, NH, Tech. Rep. TR2004-515, 2004.
Signal Process., vol. 53, no. 2, pt. 2, pp. 845–850, Feb. 2005. [22] S. Bayram, H. T. Sencar, and N. Memon, “An efficient and robust
[2] H. Farid, “Photo fakery and forensics,” Adv. Comput., vol. 77, pp. 1–55, method for detecting copy-move forgery,” in Proc. IEEE ICASSP,
2009. Washington, DC, 2009.
[3] J. A. Redi, W. Taktak, and J.-L. Dugelay, “Digital image forensics: A [23] A. C. Popescu and H. Farid, “Exposing digital forgeries by detecting
booklet for beginners,” Multimedia Tools Applicat., vol. 51, no. 1, pp. traces of resampling,” IEEE Trans. Signal Process., vol. 53, no. 2, pt.
133–162, 2011. 2, pp. 758–767, Feb. 2005.
[4] I. J. Cox, M. L. Miller, and J. A. Bloom, Digital Watermarking. San [24] S. Bayram, H. T. Sencar, and N. Memon, “A survey of copy-move
Francisco, CA: Morgan Kaufmann, 2002. forgery detection techniques,” in Proc. IEEE Western New York Image
[5] M. Barni and F. Bartolini, Watermarking Systems Engineering: En- Processing Workshop, Rochester, NY, 2008.
abling Digital Assets Security and Other Applications. New York: [25] S.-J. Ryu, M.-J. Lee, and H.-K. Lee, “Detection of copy-rotate-move
Marcel Dekker, 2004. forgery using zernike moments,” in Proc. Int. Workshop on Information
[6] H. Farid, “A survey of image forgery detection,” IEEE Signal Process. Hiding, Calgary, Canada, 2010.
Mag., vol. 2, no. 26, pp. 16–25, 2009. [26] H.-J. Lin, C.-W. Wang, and Y.-T. Kao, “Fast copy-move forgery detec-
[7] A. Popescu and H. Farid, “Statistical tools for digital forensics,” in tion,” WSEAS Trans. Signal Process., vol. 5, no. 5, pp. 188–197, 2009.
Proc. Int. Workshop Information Hiding, Toronto, Canada, 2005. [27] S. Bravo-Solorio and A. K. Nandi, “Passive method for detecting du-
[8] A. Swaminathan, M. Wu, and K. Liu, “Digital image forensics via in- plicated regions affected by reflection, rotation and scaling,” in Proc.
trinsic fingerprints,” IEEE Trans. Inf. Forensics Security, vol. 3, no. 1, EUSIPCO, Glasgow, Scotland, 2009.
pp. 101–117, Mar. 2008. [28] V. Christlein, C. Riess, and E. Angelopoulou, “A study on features for
[9] M. Chen, J. Fridrich, M. Goljan, and J. Lukas, “Determining image the detection of copy-move forgeries,” in Proc. Information Security
origin and integrity using sensor noise,” IEEE Trans. Inf. Forensics Solutions Europe, Berlin, Germany, 2010.
Security, vol. 3, no. 1, pp. 74–90, Mar. 2008. [29] X. Shuai, C. Zhang, and P. Hao, “Fingerprint indexing based on com-
[10] N. Khanna, G. T.-C. Chiu, J. P. Allebach, and E. J. Delp, “Forensic posite set of reduced SIFT features,” in Proc. ICPR, Tampa, FL, 2008.
techniques for classifying scanner, computer generated and digital [30] H. Su, A. Bouridane, and M. Gueham, “Local image features for
camera images,” in Proc. IEEE ICASSP, Las Vegas, NV, 2008. shoeprint image retrieval,” in Proc. BMVC, Warwick, UK, 2007.
[11] R. Caldelli, I. Amerini, and F. Picchioni, “A DFT-based analysis to [31] H. Huang, W. Guo, and Y. Zhang, “Detection of copy-move forgery
discern between camera and scanned images,” Int. J. Digital Crime in digital images using SIFT algorithm,” in Proc. IEEE Pacific-Asia
Forensics, vol. 2, no. 1, pp. 21–29, 2010. Workshop on Computational Intell. and Industrial Application, Wuhan,
[12] I. Amerini, L. Ballan, R. Caldelli, A. Del Bimbo, and G. Serra, “Geo- China, 2008.
metric tampering estimation by means of a SIFT-based forensic anal- [32] X. Pan and S. Lyu, “Detecting image region duplication using SIFT
ysis,” in Proc. IEEE ICASSP, Dallas, TX, 2010. features,” in Proc. IEEE ICASSP, Dallas, TX, 2010.
1110 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 6, NO. 3, SEPTEMBER 2011

[33] K. Mikolajczyk and C. Schmid, “A performance evaluation of local Roberto Caldelli (M’10) received the Laurea degree
descriptors,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 27, no. 10, in electronic engineering in 1997 and the Ph.D. de-
pp. 1615–1630, Oct. 2005. gree in computer science and telecommunication in
[34] K. Mikolajczyk, T. Tuytelaars, C. Schmid, A. Zisserman, J. Matas, F. 2001, both from the University of Florence, Florence,
Schaffalitzky, T. Kadir, and L. Van Gool, “A comparison of affine re- Italy.
gion detectors,” Int. J. Comput. Vision, vol. 65, no. 1/2, pp. 43–72, Currently he is an assistant professor at the Media
2005. Integration and Communication Center of the Uni-
[35] W. Wei, S. Wang, X. Zhang, and Z. Tang, “Estimation of image rotation versity of Florence. He is also a member of CNIT. His
angle using interpolation-related spectral signatures with application to main research activities, witnessed by several publi-
blind detection of image forgery,” IEEE Trans. Inf. Forensics Security, cations, include digital image processing, image and
vol. 5, no. 3, pp. 507–517, Sep. 2010. video digital watermarking, and multimedia foren-
[36] W. Lu, A. L. Varna, and M. Wu, “Forensic hash for multimedia infor- sics.
mation,” in Proc. SPIE Media Forensics and Security, San Jose, CA,
2010.
[37] T. Hastie, R. Tibshirani, and J. H. Friedman, The Elements of Statistical
Learning. New York: Springer, 2003. Alberto Del Bimbo (M’90) is a full professor of
[38] R. I. Hartley and A. Zisserman, Multiple View Geometry in Computer computer engineering at the University of Florence,
Vision. Cambridge, U.K.: Cambridge Univ. Press, 2004. Florence, Italy, where he is also the director of the
[39] M. Fischler and R. Bolles, “Random sample consensus: A paradigm Master in Multimedia, and the director of the Media
for model fitting with applications to image analysis and automated Integration and Communication Center. His research
cartography,” Commun. ACM, vol. 24, no. 6, pp. 381–395, 1981. interests include pattern recognition, multimedia
[40] T.-T. Ng, S.-F. Chang, J. Hsu, and M. Pepeljugoski, Columbia Pho- information retrieval, and human–computer interac-
tographic Images and Photorealistic Computer Graphics Dataset AD- tion. He has published more than 250 publications
VENT, Columbia Univ., New York, Tech. Rep., 2004. in some of the most distinguished scientific journals
and international conferences, and is the author of
Irene Amerini received the Laurea degree in com- the monograph Visual Information Retrieval.
puter engineering in 2006 and the Ph.D. degree in Dr. Del Bimbo is an IAPR fellow and Associate Editor of Multimedia
computer engineering, multimedia, and telecommu- Tools and Applications, Pattern Analysis and Applications, Journal of Visual
nication in 2011, both from the University of Flo- Languages and Computing, and the International Journal of Image and
rence, Italy. Video Processing, and was an Associate Editor of Pattern Recognition, IEEE
She was a visiting scholar at Binghamton Uni- TRANSACTIONS ON MULTIMEDIA, and IEEE TRANSACTIONS ON PATTERN
versity, Binghamton, NY, in 2010. Currently she ANALYSIS AND MACHINE INTELLIGENCE.
is a postdoctoral researcher at the Image and
Communication Laboratory, Media Integration and
Communication Center, University of Florence,
Florence, Italy. Her main research interests focus on
Giuseppe Serra received the Laurea degree in com-
multimedia forensics and image processing.
puter engineering in 2006, and the Ph.D. degree in
computer engineering, multimedia, and telecommu-
nication in 2010, both from the University of Flo-
rence, Florence, Italy.
Lamberto Ballan (S’08) received the Laurea degree He is a postdoctoral researcher at the Media
in computer engineering in 2006 and the Ph.D. degree Integration and Communication Center, University
in computer engineering, multimedia, and telecom- of Florence. He was a visiting scholar at Carnegie
munication in 2011, both from the University of Flo- Mellon University, Pittsburgh, PA, and at Télécom
rence, Italy. ParisTech/ENST, Paris, in 2006 and 2010, respec-
He was a visiting scholar at Télécom ParisTech/ tively. His research interests include image and
ENST, Paris, in 2010. Currently he is a postdoctoral video analysis, multimedia ontologies, image forensics, and multiple-view
researcher at the Media Integration and Communica- geometry.
tion Center, University of Florence, Florence, Italy.
His main research interests focus on multimedia in-
formation retrieval, image and video analysis, pattern
recognition, and computer vision.