2015 International Symposiwn on Advanced Computing and Communication (ISACC)

Digital Image forgery detection using SIFTfeature

Rajeev Rajkumar Kh. Manglem Singh
Department of Information Technology Department of Computer Science & Engg,.
Assam University Silchar, Assam National Institute of Technology, Manipur
[email protected] [email protected]

Abstract- There has been an extensive growth in the area of

digital image processing. One of the main issues in the real world
is to judge the authenticity of a given image. Nowadays it is very
easy to tamper and forge digital image with the advancement of
the powerful digital image processing software and easy
availability of the tools. The most common form of image
manipulation techniques is the region duplication also called as
copy move forgery where a portion of the image is copied and
paste to another portion in the same digital image. To investigate
such forensic analysis, various techniques and method have been
developed in the past literature. There has been an immense
research of finding such tampered pixel blocks. This method
cannot handle cases when the copied region is scaled or rotated to
a new region. This paper proposes an efficient algorithm based
on scale invariant feature transform (SIFT) with feature
extraction which is invariant to translation, scale, noise and
rotation. It comprises transformation of the input image to
produce a standard representation and then detection of
keypoint and feature descriptor is applied along with a matching
over all the keypoints.

Keywords-SIFT; Difference of Gaussian; Copymove forgery,

composite image, tampered image

I. INTRODUCTION

Nowadays lots of sophisticated powerful digital cameras and

photo-editing software tools are freely available. It has
become very easy to alter digital images and create tampered (c)
images that are very difficult to distinguish from the original Fig 1: Different types of image forgery (a) Tampered Image
authentic images. With the development of new advance (b) Composite Tmage (c) Copymove Image
technology, availability of fast and powerful computing tools
such as Adobe Photoshop and coreldraw etc, it is very easy to
Despite of this fact, detection all types of tampering accurately
manipulate, forge or tamper digital image without leaving any
is not available till now. So, there is tremendous increase of
obvious clue. An image can be tampered in various ways:
the digitally manipulated forgeries in the digital media and on
deleting or hiding a segment in the image, adding a new object
the Internet. Therefore it is very much important to develop
in the image and misrepresentation of image information.
techniques that can verify the integrity and authenticity of
In our daily life, we see things which are not always what we particularly images in applications such as forensic
think they usually look like. It is just because we believe investigation, insurance processing, surveillance systems,
something to be true does not necessarily means it is true. It is intelligence services, medical imaging, journalism etc. Digital
also just because we do not believe something does not mean images are also used in the court as evidence of crime, and are
that it is not true. Authenticity is the basic requirement to also used by physicians.
believe the originality of the data, which may be image or
Image tampering as shown in Fig.1 can be broadly classified
video.
into categories: tampering image feature, image composition
and copy-move forgery. In copy-move forgery, a single image
is used to perform forgery within the image. The most
common manipulation in tampering with a digital image is to
copy and paste portions of the image to conceal a person or
object in the scene. A copy-move forgery is a specific type of
digital image manipulation technique where a portion of the
image is being copied and pasted to another portion within the

978-1-4673-6708-0/15/$31.00 ©2015 IEEE

same image. Usually motivations are either to hide or to II. PREVIOUS WORK IN FORGERY DETECTION
conceal unwanted portion of an image, or to focus on
Detection of forgery in digital image can be divided into two
particular portion the image. The copied and the pasted region
type: active and passive (blind) method. Active approach
are from the same digital image which includes properties like
includes digital watermarking and steganography, Digital
noise, illumination, rotation are expected to be matched on the
watermarking is used for a fragile authentication, owner
tampered region of the digital image Most techniques follow a
identification, content protection, detection of tampering,
common pipeline, as shown in Fig. 2. Copy-move forgery
localization of changes, copyright protection and recovery of
detection methods are either key point-based methods [3,11],
original content [4,5]. The main disadvantage of active
or block-based methods [4,7,9,lO,12-[18]. Preprocessing of
approach is that the signature must be present in the digital
the images is possible in both the two cases. However, for
image prior to the tampering occurs, and most digital cameras
instance most methods operate on gray- scale images, and as
and digital devices do not have such instantaneous embedding
such there is requirement to merge the color channels.
facilities [6]. So, it is not always possible to embed signature
Copy-move forgery detection methods are either keypoint­ or any digital form of information in the digital image by the
based methods or block-based methods. In both method, it is digital devices to prove authenticity.
possible to pre-process the given digital images. As of now,
The limitations of active approach has led to the development
most of the methods are perform on gray- scale images. In
of the passive method which exploits various kind of intrinsic
feature extraction part, block-based methods sub-divide the
fmgerprint like sensor noise of the capturing device or
digital image into smaller block. A feature vector is computed detectable changes in image for detecting forgery [7]. The
for every such block. Similar feature vectors of every blocks
blind approach can detect textures which are occur in the real
are subsequently matched.
image using different techniques and find the inconsistency in
those techniques between different areas of the image.
However, it may produce a large number of false positives and
On the other hand, without any image subdivision keypoint­ missed detections.
based methods compute the features only on image regions
with high entropy. Then the Similar features of the image are
matched. If the regions of such matches cluster into larger III. SURVEY ON COPY-MOVE FORGERY DETECCTION
areas, then forgery are detected. In both methods, it include For copy-move forgery, a portion of the digital image is
further post filtering operation for removing false matches. In duplicated and pasted to different part of the same digital
order to group matches that jointly follow a transformation image that are performed with the intention to conceal an
pattern, other optional postprocessing step of the detected object from the image by putting over it with a duplicate part
regions of the image may be performed. from the another area of the same digital image. This
duplicated part is usually an object like gravel, grass or texture
with irregular patterns. Such region are perfect environment
for image forgery since the duplicated area will simply blend
with the background and they are not be able to detect any
doubtful artifacts easily by human eye. Noise component,
color palette, dynamic range and other important properties of
the tampered image are compatible with the original image, as
copied part comes from the same image.
Fridrich et al proposed the exact method in copy-move forgery
detection [8]. Their method detects the exact and same copy of
a region by sliding a square block of pixels across the image
from the upper left corner down to the lower right corner in a
step of one pixel to obtain overlapping blocks. It is hard to
detect in sophisticated manipulation. The proposed method
matches based on comparison of quantized discrete cosine
transform (DCT) coefficients. By varying the quantization
factor, the degree of robustness can be controlled. More false
matches can be obtained with the higher quantization factor.

Fig. 2: Common forgeIY techniqes for the detection of copy-move forgeries.

Popescu and Farid proposed a method based on principal
The feature extraction differs for keypoint-based features (top) and block-based component analysis (PCA) instead of DCT representation [9].
features (bottom). The accuracy of their method is good except for small block
size and low JPEG qualities.
Section II gives a previous work done in image forgery Huang and Qiu proposed a copy-move forgery detection
detection. Section III describes different methods of copy­ which has lower computational complexity and robust against
move forgery detection. Section IV gives the detailed SIFT blurring, noise contamination, severe lossy compression, and a
feature followed by Experimental Result in Section V. mixture of these processing operations POl Langille and
Conclusion is given in Section 6. Gong proposed a technique of detecting duplicate region that
operate Zero mean Normalized Cross Correlation (ZNCC) Solario and Nandi proposed a method that uses log polar
rI11, Their method is not robust against rotation and scaling coordinates [24]. The method operates on reassembling the
attacks. overlapping block into their log polar coordinates and they are
summed along the angle axis for obtaining a one dimensional
Li et al proposed a copy move forgery detection which
descriptor which are invariant to reflection and rotation.
reduces the block size that uses Discrete Wavelet Transform
Correlation coefficient of its Fourier magnitudes is used for
(DWT) and Singular Value Decomposition (SVD) [13] with
performing more effecting search of duplicate region.
improvement in time complexity of the detection algorithm.
The author does not reveal the robustness against rotation, Lin et al proposed a method using radix sort to increase the
noise, illumination and scaling of the copied region in spite of efficiency [25]. The image is spliced into overlapping blocks
claiming of high accuracy in the presence of JPEG which are of same size. Their method is robust against
compression. Myna et al proposed another method based on compression and Gaussian noise. Wang et al proposed a
DWT and SVD, which converts spatial coordinates to log method based on Gaussian pyramid [26]. Their method could
polar coordinates to oppose the effect of scaling and rotation not resist against random rotation.
[14]. Their method is robust against scaling and rotation.
Junhong proposed a new method which is based on Locally
Feng and Su proposed methods which are based on DWT and Linear Embedding (LLE) [27]. LLE are usually mean for
SVD to lower the dimension. But block matching was not reducing the dimensionality in high dimensional data set. The
used. To estimate the spatial offset between copied region and method is invariant to rotation, horizontal flips, vertical flips,
pasted region, phase correlation was computed instead of translation, but the computational cost is expensive.
using block matching. The copy move regions were
highlighted by pixel matching. For compressed images, it can IV. SCALE INVARIANT FEATURE TRASFORM
also be detected [15].
David Lowe proposed Scale Invariant Feature
Mahdian and Saic proposed a technique which are based on Transform which is a digital image descriptor for matching
blur moment invariants [16,17]. The method concentrates on and recognition of digital image. The SIFT descriptors are of
detecting copied region with the presence of blurring and 128 descriptor vectors which are mainly used in point
adding noise. The main idea behind them is the characteristic matching between different views of a 3-D scene and object
of blur moments that small scale blurring will not affect main recognition in computer vision. These descriptor are robust to
blur moments. rotation, scaling, translation transformation in the image and
Dybala et al proposed a method for copy move forgery also robust to illumination variations. Therefore SIFT are
detection by detecting the use of filters that are used to useful for image matcing and recognition under real world
thinning the pasted region [18]. Their method shows environment.
robustness to high quality compression. The SIFT descriptor consist of detecting interest point from
Lin et al proposed a technique for copy move forgery the image which is grey level image for which local gradient
detection in JPEG images. The method can also be detected direction of the image intensities are put together to give a
both for an alteration performed on JPEG and the target JPEG brief description of the image structure in a local
image [19]. Their method is robust against JPEG compression, neighbourhood around the keypoint, which should be used for
but it is hard to detect, if image is compressed thrice. matching corresponding keypoints between the different
image. Then the SIFT descriptor is applied at the dense grids
Bovik and Liu proposed a method based upon block artifact which are shown to cause the better performance for object
(BA) [20]. Their method works for multi compressed images. categorization, texture classification, image alignment and
biometrics. The SIFT descriptor can also be used for colour
Lowe proposed a new technique which is based on Scale
image and 2+ I-D spatio-temporal video. An overview of the
invariant Feature Transform algorithm (SIFT) [21]. It is
algorithm is presented here. There are mainly four major
performed by extracting 128 vector SIFT descriptors of the
stages of computation involved in SIFT algorithm.
given digital image. The SIFT descriptors are invariant to
changes in illumination, rotation, scaling etc which is the main
criteria of their technique. For detecting the tampered region, A. Scaled-space Extrema Detection
SIFT descriptors of copied and pasted region are computed Scaled-space Extrema Detection is the first steps that searches
and matched. The method computes the distinctive features of over all scales and image locations. It uses a difference-of­
local image patches that are invariant to scaling and rotation Gaussian function for identification of potential interest points
and that also are robust to changes in noise, illumination etc. which are invariant to scale and orientation. Laplacian of
Their method is robust against JPEG compression, but it gives Gaussian (LoG) is calculated for the image with various (J
more false positive for small block size. values which acts as a blob detector and detect blobs of
different sizes with the change in (J. Gaussian kernel with low
Bayram et al proposed a method which is based on Fourier­ (J will give a high value for small corner in an image while
Mellin Transform (FMT) for computing features of the image guassian kernel with high (J fits well for broader corner. Thus,
block of the tampered image [22]. For comparing the similar it will fmd the local maxima across the different scale and
blocks of the image, lexicographic sorting is used. Their space which will again give a vector of (x,y,(J) values. These
method can also be well detected in the presence of JPEG mean that at (J scale there is a potential keypoint at (x,y) which
compression, blurring, noise addition, rotation, and scaling. are shown in equation 1.
 . (1) Normally, the value of different parameter that are used in this
technique are given as number of different octaves = 4,
number of scale levels value = 5, initial σ value =1.6, etc. as
where is the blurred image with amount of blur, feasible values..

is the Gaussian Blur operator ,

is the pixel at row and column of the image
, and is the two dimensional convolution operator in and
. If the amount of blur in a particular image is σ, then the
amount of blur in next level will be , is a constant,

which is given by . LoG is little

costly compared to Difference of Gaussians, therefore SIFT uses
DoG that are the approximation of LoG. DoG are computed from the
Fig 4: Local Extrema over scale and space
difference of Gaussian blurring of an image with two adjacent value
of σ, let it be σ and kσ as shown in equation 2 B. Keypoint Localization
(2)
At each candidate location, a detailed model is fit to determine
location and scale. Keypoints are selected based on measures of their
where and are the blurred image stability. After the potential keypoint location is calculated, it has to
with blur amount and respectively be refined to obtain more accurate results. For that Taylor series
expansion of scale space is used to obtain more correct location of
For different octaves of the image, this process is performed in extrema. The keypoint is rejected if the intensity of this extrema ies
Gaussian Pyramid as shown ion figure 3.. less than a predefined threshold value (normally 0.003). Edges also
need to be removed as DoG has higher response for edges which uses
Harris corner detector is used. For computation of principle
curvature, it uses a 2x2 Hessian matrix (H). For every candidate
key-point at coordinate , the Hessian matrix is
calculated as follows:

(3)

where , , and are

Fig 3: Difference of Gaussian (DOG)

Then the local extrema over scale and space are searched on
the image. For example one pixel value is compared to its 8
neighbours as well as 9 pixels in next scale and 9 pixels in If , then retain the key-point,
previous scales and so on.. For every pixel with spatial otherwise discard it. where Cedge is the ratio between the largest
location and in image , the pixel with spatial and non-zero smallest eigen-values in the block of the image. From
location Harris corner it is known that for edges, one value is larger than the
other. The keypoint is rejected if this ratio is higher than a threshold.
Usually it uses 10 as a threshold. This step eliminates the low
contrast keypoint and edge keypoint and the only accurate keypoint is
obtained.
in the current image and consecutive scale images and
C. Orientation Assignment
are the neighbouring pixels along with positions in Based on local image gradient directions one or more orientations are
and . If any one of the pixel value is the greatest of all assigned to each keypoint location. All the image operation that has
the 26 neighbours, then the pixel is considered as the maxima been transformed are performed relative to the assigned orientation,
scale, and location for each feature which provide invariance to these
point, and if any of the pixel value is the least of all the transformations. In order to achieve invariance to image rotation an
neighbours, the pixel is minima point. All these maxima and orientation is assigned to each keypoint. Depending on the scale, a
minima points are considered as candidate keys. Potential neighbourhood point is chosen around the keypoint location. Then for
keypoint is identified if it is a local extrema which basically that region the gradient magnitude and direction is calculated. After
means that keypoint is best represented in that scale which is that it is the creation of orientation histogram with 36 bins covering
shown in figure 4: 360 degrees. It is usually weighted by gradient magnitude and
gaussian-weighted circular window with (J equal to 1.5 times the
scale of keypoint. Then in the histogram the highest peak is chosen
and any peak above 80% of it is also taken for calculating the
orientation which creates a keypoint with same location and scale.
They also contribute to the stability of matching.

D. Keypoint Descriptor

At the selected scale, the local image gradients are measured

in the region around each keypoint. They are converted into a
representation that also works for significant levels of local
shape distortion and illumination.
Keypoint descriptors are now obtained. A 16x16
neighbourhood around the keypoint is selected and divided
into 16 sub-blocks of 4x4 size block. For every sub-block, 8 (a)
bin orientation histogram is created which lead to a 128 bin
value which also represents a vector forming keypoint
descriptor. There are several techniques which are meant to
obtain robustness to illumination, rotation, noise etc.

E. Keypoint Matching

By identifying their nearest neighbours, keypoint between

similar images are matched. It can also be happen that the
second closest-match may be very near to the first which may
due to noise and some other factors. For this case, ration of
closest-distance to second -closest distance is obtained. They
are rejected if it is greater than 0.8. These steps discarded
almost 90% of false matches and nearly only 5% correct
matches are discarded.. (b)
Fig 5: Original image of (a) Lena and (b) man
V. EXPERIMENTAL RESULT

The SIFT keypoint of the original images as shown in Fig. 5 is

calculated which is shown in Fig.6. In the former image of
lena the left eye of the lena was copied and pasted on the right
side, and the nose is scaled and rotate to 90 was again pasted
to the right side of the image and in the latter image the car is
scaled and copied and moved onto the road area. The proposed
detection method is used to detect tampered regions in them.
SIFT keypoints features of the two images as shown in Fig.7
are extracted by first, and with a threshold their descriptors are
matched within each other. A greater threshold leads to more
match points which then tend to come with more false
matches, while a smaller value will get more accurate matches
with lesser match points. For that, it is set to be 0.30, to strike
a balance between numbers of correct matches and total
numbers of matches. The detection results are displayed in the
image with shaded region as shown in Fig. 8. By observing the
detection results, the duplication regions can be figured out.

(b)
Fig 6: Forged image of (a) Lena and (b) man
 [3] e. Zhang, L.L. Cheng, Q. Zhengding and L.M. Cheng, "Multipurpose
watermarking based on multiscale curvelet transform",IEEE Trans. Inf
Forensics Security,vol. 3,no. 4,pp. 611-619,2008.
[4] J. Fridrich, "Method for tamper detection in digital images", Proc.
ACM Workshop on Multimedia and Security,pp. 19-23,1999.
[5] 1. T. Hsieh and Y.K. Wu, "Geometric invariant semi fragile image
watermarking using real symmetric matrix", WSEAS Trans. of Signal
Processing,vol. 2,no. 5,pp. 612-618,2006.
[6] H. Farid,"image forgery detection - a survey",IEEE Signal Processing
Magazine,vol. 5,pp. 16-25,2009.
[7] G. Muhammad, M.H. AI-Hammadi,M. Hussain, and G. Bebis, "Image
forgery detection using steerable pyramid transform and local binary
pattern ", Machine Vision and Applications, DO[ 10. 1007/s00138-013-
0547-4,2013.
[8] J. Fridrich,D. Soukal and 1. Lukas,"Detection of copy-move forgery in
digital images", Proc. IEEE Digital Forensic Research Workshop, pp.
55-61,2003.
[9] Ae. Popescu and H. Farid, "Exposing digital forgeries by detecting
traces of resampling," IEEE Trans. on Signal Processing,vol. 53(2), pp.
(a) (b) 758-767,2005.
Fig 7: SIFT Keypoint of (a) Lena and (b) man [10] W. Luo, 1. Huang and G. Qiu, "Robust detection of region-duplication
forgery",Proc. IEEE ICPR,pp. 746-749,2006.
[II] A Langille and M. Gong, "An efficient match-based duplication
detection algorithm",IEEE CRV,p. 64,2006.
[12] I.K. Jung and S. Lacroix, "A Robust Interest Point Matching
Algorithm",Proc. [nternational Conference on Computer Vision. 2001.
[13] G. Li, Q. Wu, D. Tu, and S. Sun, "A Sorted Neighborhood Approach
for Detecting Duplicated Regions in Image Forgeries based on DWT
and SVD," Proc. IEEE ICME,2-5,pp. 1750- 1753,2007.
[14] AN. Myna, M.G. Venkateshmurthy and e.G. Patil, "Detection of
region duplication forgery in digital images using wavelets and log­
polar Mapping",Proc. IEEE ICCTMA, pp. 371-377,2007.
[15] 1. Zhang, Z. Feng, and Y. Su, "A new approach for detecting copy­
(a) (b) move forgery in digital images",Proc. IEEE ICCS,pp. 362 -366,2008.
Fig 8: Result showing the copied region of (a) lena (b) man [16] B. Mahdian and S. Saic, "Detection of copy-move forgery using a
method based on blur moment invariants", Forensic Science
International, 171 (2-3),pp. 180-189,2007.
ACKNOWLEDGMENT [17] J. Flusser, T. Suk and S. Saic,"Image features invariant with respect to
We would like to acknowledge the staffs of Department of blur",Pattern Recogn. 28 ,pp. 1723-1732,1995.
Information Technology, Assam University for their support [18] B. Dybala, B. Jennings and D. Letscher, "Detecting filtered cloning in
and for their guidance and constant encouragement towards digital images", Proc ACM MM&Sec Workshop,pp. 43-50, 2007.
[19] J.F. He, Z.C. Lin, L.F. Wang, and X.O. Tang, "Detecting doctored
the work.
JPEG images via DCT coefficient analysis", Lecture Notes in
CONCLUSION Computer Science,Springer Berlin,vol. 3953,pp.423-435,2006.
[20] Ae. Bovik,and S. Liu,"DCT-domain Blind Measurement of Blocking
A SIFT based forgery detection in digital image is used in this Artifacts in DCT-coded [mages", Proc. IEEE [CASSP, pp. 1725-1728,
paper. This SIFT based technique which are dependent on 2001.
[21] D.G. Lowe, . "Distinctive Image Features from Scale-Invariant Key
feature extraction using keypoint and feature descriptor which points". International Journal of Computer Vision, vol. 60, no. 2, pp.
are invariant to scaling, noise, illumination and rotation. It is 91-110,2004.
then performed matching and clustering to every keypoints [22] S. Bayram, H. Taha Sencar and N. Memon, " An efficient and robust
and matching points for detecting tampered region. This method for detecting copy-move forgery", Proc IEEE ICASSP, pp.
method is commonly used for detecting SUSpICIOUS 1053-1056,2009.
[23] J.F. He, Z.C. Lin, L.F. Wang, and X.O. Tang, "Detecting doctored
manipulation with digital images (digital forgeries).
JPEG images via DCT coefficient analysis", Lecture Notes in
Computer Science,Springer Berlin,vol. 3953,pp.423-435,2006.
REFERENCES [24] S.B. Solorio and AK. Nandi, " Passive forensic method for detecting
duplicated regions affected by reflection, rotation and scaling" , Proc.
[I] M.M. Yeung, "Digital watermarking", ACM Commun., vol. 41, no. 7, European Signal Proc. Conf. 2009,pp. 824-828.
pp. 30-33, 1998. [25] H. J. Lin, C.-W. Wang, Y.-T. Kao, Fast copy-move forgery detection,
[2] e. Rey and J.L. Dugelay, "A survey of watermarking algorithms for WSEAS Transactions on Signal Processing,5 (5),pp. 188-197,2009.
image authentication",EURAS[P Journal on applied Signal Processing, [26] J.Wang, G. Liu, H. Li, Y. Dai and Z. Wang, "Detection of [mage
vol. 2002 N6,pp. 613-62[,2002. Region Duplication Forgery Using Model with Circle Block",
International Conference on Multimedia Information Networking and
Security,vol. I, pp.25-29,2009