16-05-2013, 11:44 AM
Exposing Postprocessed Copy–Paste Forgeries Through Transform-Invariant Features
Exposing Postprocessed.pdf (Size: 1.77 MB / Downloads: 31)
Abstract
Image manipulation has become commonplace with
growing easy access to powerful computing abilities. One of the
most common types of image forgeries is the copy–paste forgery,
wherein a region from an image is replaced with another region
from the same image. Most prior approaches to finding identical
regions suffer from their inability to detect the cloned region when
it has been subjected to a geometric transformation. In this paper,
we propose a novel technique based on transform-invariant features.
These are obtained by using the features from the MPEG-7
image signature tools. Results are provided which show the efficacy
of this technique in detecting copy–paste forgeries, with translation,
scaling, rotation, flipping, lossy compression, noise addition
and blurring. We obtain a feature matching accuracy in excess of
90% across postprocessing operations and are able to detect the
cloned regions with a high true positive rate and lower false positive
rate than the state of the art.
INTRODUCTION
TAMPERING images has become extremely easy due to
the easy accessibility of advanced image editing software
and powerful computing hardware. Various types of forgeries
can be created and in recent years, image forgery detection using
passive techniques has become a hot area of research [1], [2].
One of the most common types of image forgeries is the
copy–paste (or copy–move or cloning) forgery, where a region
from one part of an image is copied and pasted onto another
part, thereby concealing the image content in the latter region.
Such concealment can be used to hide an undesired object or
increase the number of objects apparently present in the image.
Although a simple translation may be sufficient in many cases,
additional operations are often performed in order to better hide
the tampering. These include scaling, rotation, lossy compression,
noise addition, blurring, among others.
PRIOR WORK IN COPY–MOVE FORGERY DETECTION
As discussed earlier, copy–move forgeries involve concealing
one region in an image by overlaying another region
from the same image. The most seemingly obvious way of
detecting copied and pasted regions in the same image would
be to verify small clusters or blocks of pixels for matches all
across the image. However, there are two major issues with this
approach. Firstly, this would be a computationally intensive
approach, as matching blocks (or other shapes) of pixels would
become infeasible with increasing size of the image. Secondly,
such an approach would fail in case of minor changes such as
addition of noise or multiple image compression. In order to
circumvent these drawbacks of this direct approach, researchers
have developed various techniques which can be classified into
two main categories: block-based and feature-based.
Block-Based Techniques
Some techniques use representations for dimensionality reduction
[3], [4] such as principal component analysis (PCA) or
frequency representation [5] such as discrete cosine transform
(DCT) in order to efficiently find matching regions. However,
they assume that the copied region has not undergone any postprocessing,
which is not always the case. Nevertheless, these
techniques are invariant to slight noise addition and lossy compression.
The work of [6] discusses improved robustness using
DCT to noise addition, global blurring and lossy compression,
but does not deal with geometrical transformations of the tampered
region. The technique of [7] reduces the time complexity
of the PCA-based approach by using a discrete wavelet transform
(DWT), but again does not address geometrical transformations.
Feature-Based Techniques
Block-based techniques essentially compare blocks in an efficient
manner and provide invariance to some transformations
through an appropriate choice of the method of representation.
It is seen, however, that this often results in significant false
positives, and invariance to other transformations like flipping,
brightness changes and blurring is hard to establish. Therefore,
recently, interest in feature-based approaches has been spurred,
as forgeries have become more convincing with a number of
transformations being employed. Feature-based techniques try
to avoid these problems by choosing to match features in the
image, instead of blocks. By an appropriate selection of features,
invariance to a number of transformations can be established.
The rationale for this lies in the fact that the features of interest
were developed for the purposes of object detection and/or
content-based image retrieval and so needed to be invariant to
a large number of transformations. Our proposed technique is
an example of this class of copy–move forgery detection techniques
as well.
PROPOSED TECHNIQUE
We propose a novel technique for detecting copy–paste forgeries
with possible postprocessing. It is based on the MPEG-7
image signature tools [16], which form a part of the MPEG-7
standard. This set of tools was designed for robust and fast
image and video retrieval. The main issue in directly applying
these tools to image forgery detection is that these tools were designed
to find duplicate, but separate, images, whereas we are
trying to find identical regions in the same image.We perform
modifications in the feature extraction and matching processes
to efficiently detect copy–paste forgeries.
Feature Matching
As the descriptor components are no longer binary, we do not
employ a Hamming distance measure, as in the original tools.
Instead, the Euclidean distance between the components of each
pair of features (denoted by ) is calculated, and the feature
pairs belowa threshold, , are chosen for further analysis. This
threshold is set sufficiently high to remove only the unlikeliest
matching feature pairs, without eliminating any true matches.
Additionally, feature pairs where the constituent features have
a spatial Euclidean distance of less than 20 pixels are ignored.
This is important because features which are located close to
each other tend to be similar due to the smoothness constraints
of natural images, without being part of cloned regions.
Comparisons
A quantitative analysis of various copy–move forgery detection
methods has been performed on a database in [22].
Examples of detection of copy–move forgeries by our method
for this database are shown in Fig. 12(a)–©. For the image
in Fig. 12©, the small falsely matched regions along the
shoreline arise because of the geometrical transformation
between the cloned regions resulting in a high correlation
coefficient along the relatively uniform shoreline. However, no
matching features are found in the falsely matched regions by
our proposed technique which indicates that the false positives
are indeed caused by the nature of the image content and are
unavoidable if the cloned regions are to be indicated. Since
our method specifically addresses copy–move forgeries that
undergo various forms of postprocessing, we compare our
method with various aspects of the state-of-the-art methods in
[14], [15] having the same capabilities. These use SIFT features
to detect copy–move forgeries.
CONCLUSION
Copy-move forgeries are a common type of forgery where
parts of an image are replaced with other parts from the same
image. The copied and pasted regions may be subjected to various
image transformations in order to conceal the tampering
better. Conventional techniques of detecting copy–paste forgeries
usually suffer from the problems of false positives and susceptibility
to many image processing operations.
In this paper, we have proposed a technique based on the
MPEG-7 image signature tools, which have been developed
for robust content-based image retrieval, in order to detect
copy–move forgeries. We have modified the tools in many
ways to deal with copied regions in a single image. We have
used a feature matching process that utilizes the inherent
constraints in matched feature pairs to improve the detection
of cloned regions.