Professional Documents
Culture Documents
Image Forgery Detection and Deep Learning Techniques: A Review
Image Forgery Detection and Deep Learning Techniques: A Review
Image Forgery Detection and Deep Learning Techniques: A Review
Abstract—Due to the easily available software for tampering of forged images, many types of research have been done to
images, image manipulation has become quite common. Since the detect manipulated images. Active approaches for forgery
tampered images are non-distinguishable by the naked eye, they detection which include digital signatures and watermarking
are being circulated on various platforms giving rise to rumors
and misleading many. This has led researchers to work on suffer fro m the drawback of inserting the watermark or the
various techniques for the detection of manipulated images with signature beforehand in the image which limits the scenarios
improved accuracy. Traditional works on image forgery detection where this technique would work. Various approaches have
are mostly based on extracting simple features that are specific been proposed for passive based forgery detection. Popescu
for detecting some particular type of forgery. Recently, works on Farid [10] described a method for detecting duplicate regions
forgery detection based on neural networks have proved to be
very efficient in detecting image forgery. Neural networks are
present in an image by apply ing PCA to blocks of images and
capable of extracting complex hidden features of an image, thus then sorting the blocks lexicographically. Amerini et al.
giving better accuracy. Contrary to the traditional methods of [1] has used a SIFT based approach that solves the two-fold
forgery detection, a deep learning model automatically builds the purpose of detecting the copy move forgery and to fetch the
required features, hence it has become the new area of research in geometric transformation applied to build the forged image.
image forgery. The paper initially discusses various types of image
forgery techniques and later on compares different approaches Wei et al. [12] proposed a rotation angle estimation method
involving neural networks to identify forged images. that can detect splicing forgery. The algorithm is capable of
Index Terms—Copy move forgery, Image splicing, blind detection of geometrical operations performed and to
Convolutional neural network, Deep neural network detect foreign areas present in t h e forged image. Ke et al.
[7] used a technique to detect forged images by determin ing
I. INT RODUCT ION the consistency of shadow. It assumes that the tampered image
The emergence and widespread usage of social media has an inconsistent shadow due to splicing. The method has a
platforms like Instagram, WhatsApp, Youtube, etc. have shortcoming that it only works for forged images with shadow.
increased the number of images being uploaded and shared on Another method to identify forgery in images with reflect ions
these platforms. These images are used to spread informat ion is developed by Farid [9] which detects geometrical
to a widespread audience and hence helps to form public inconsistencies based on 3-Dimensional geo metry, reflection,
opinion on a large scale. Due to this underlying property of and projection rules. Most of these methods involve the use of
forming mass opinions and easily available photo and video structural or geo metrical properties of images to ext ract
editing tools, many false images are circulated on these features. Image forgery detection methods that involve the use
platforms daily. Not only on social media, but tampered of deep learning techniques have become very popular
images are also used in courtrooms, scientific journals, nowadays. These approaches do not consider any statistical or
literature works, etc. Image forgery is a term that refers to geometrical calculations but tend to construct a model that
man ipulating or tampering the original image to hide some automatically ext racts useful features to perform the
useful information or to showcase some false information. The classification. The use of convolution neural networks has
purpose behind creating forged images could range from shown very good results over image manipulation detection.
earning money, spreading rumors, or making false claims in The paper presents an overview of basic concepts of image
one’s favor. Due to t h e widespread use of images and forgery and various research methodologies involving deep
serious consequences learning adopted so
Authorized licensed use limited to: University of Exeter. Downloaded on June 20,2020 at 08:03:09 UTC from IEEE Xplore. Restrictions apply.
Proceedings of the International Conference on Intelligent Computing and Control Systems (ICICCS 2020)
IEEE Xplore Part Number:CFP20K74-ART; ISBN: 978-1-7281-4876-2
far for image forgery detection. The paper is div ided into the image sources. However, more than two image sources can
following sections. Section 2 discusses various techniques of also exist.
image forgery detection. Section 3 presents an overview and
general structure of the Convolutional Neural Network (CNN).
Various research approaches for manipulat ion detection using
CNN are co mpared and contrasted in section 4. The
conclusion of the paper is presented in Section 5.
A. Active approach
In the active approach of fo rgery detection, the image is
pre-processed and some cipher key is inserted in the image.
The received image is authenticated by the use of that key
[11]. So me examp les of act ive approaches include digital
watermarking and signatures.
B. Passive approach
In this type of forgery detection method, no pre-processing
of the image is required. It is based on the fact that tampering
the original image results in producing inconsistencies within
Fig. 2. Image Splicing
its statistical properties or pixel intensities. These
inconsistencies are detected to find whether the image has
tampered or not. This technique is more popular than an active Copy move forgery
approach because of the non-requirement of any prior
As the name itself indicates, this type of forgery involves
informat ion of the image. This is further classified into two
copying one part of image and pasting it in some other
types-forgery independent and forgery dependent. Forgery location. The purpose is main ly to hide some important
dependent methods detect a specific type of forgeries whereas information in the image.
forgery independent methods detect forgery independent of
their type. 2) Forgery independent: Image retouching
1) Forgery dependent: These techniques are classified into This is the most widely used and most easy to do form
copy-move and image splicing. of image tampering wherein the features of the image are
Image splicing changed in order to produce tampered image. Th is may involve
Splicing refers to cutting out a part of one image and putting rotation, resizing, changing the brightness, sharpness, color
it in the other, thus the tampered image is produced by a contrast etc of the image.
combination of two images. The produced image has two
Authorized licensed use limited to: University of Exeter. Downloaded on June 20,2020 at 08:03:09 UTC from IEEE Xplore. Restrictions apply.
Proceedings of the International Conference on Intelligent Computing and Control Systems (ICICCS 2020)
IEEE Xplore Part Number:CFP20K74-ART; ISBN: 978-1-7281-4876-2
Authorized licensed use limited to: University of Exeter. Downloaded on June 20,2020 at 08:03:09 UTC from IEEE Xplore. Restrictions apply.
Proceedings of the International Conference on Intelligent Computing and Control Systems (ICICCS 2020)
IEEE Xplore Part Number:CFP20K74-ART; ISBN: 978-1-7281-4876-2
T ABLE I
Authorized licensed use limited to: University of Exeter. Downloaded on June 20,2020 at 08:03:09 UTC from IEEE Xplore. Restrictions apply.
Proceedings of the International Conference on Intelligent Computing and Control Systems (ICICCS 2020)
IEEE Xplore Part Number:CFP20K74-ART; ISBN: 978-1-7281-4876-2
Authorized licensed use limited to: University of Exeter. Downloaded on June 20,2020 at 08:03:09 UTC from IEEE Xplore. Restrictions apply.