Central Washington University

ScholarWorks@CWU

All Master's Theses Master's Theses

Spring 2020

Image Forgery Detection with Machine Learning

Lubna Alzamil
[email protected]

Follow this and additional works at: https://digitalcommons.cwu.edu/etd

Part of the Computer Engineering Commons

Recommended Citation
Alzamil, Lubna, "Image Forgery Detection with Machine Learning" (2020). All Master's Theses. 1361.
https://digitalcommons.cwu.edu/etd/1361

This Thesis is brought to you for free and open access by the Master's Theses at ScholarWorks@CWU. It has been
accepted for inclusion in All Master's Theses by an authorized administrator of ScholarWorks@CWU. For more
information, please contact [email protected].
IMAGE FORGERY DETECTION WITH MACHINE LEARNING

A Thesis
Presented to
The Graduate Faculty
Central Washington University

In Partial Fulfillment
of the Requirements for the Degree
Master of Science
Computational Science

by
Lubna Alzamil
June 2020
 CENTRAL WASHINGTON UNIVERSITY
Graduate Studies

We hereby approve the thesis of

Lubna Alzamil

Candidate for the degree of Master of Science

APPROVED FOR THE GRADUATE FACULTY

Dr. Razvan Andonie

Dr. Boris Kovalerchuk

Dr. Szilard Vajda

Dean of Graduate Studies

ii
 ABSTRACT

IMAGE FORGERY DETECTION WITH MACHINE LEARNING

by

Lubna Alzamil

June 2020

The issue of forged images is currently a global issue that spreads mainly via social

networks. Image forgery has weakened Internet users confidence in digital images. In

recent years, extensive research has been devoted to the development of new techniques to

combat various image forgery attacks. Detecting fake images prevents counterfeit photos

from being used to deceive or cause harm to others. In this thesis, we propose methods

using the error level analysis algorithm to detect manipulated images. We show that our

combination of image pre-processing and machine learning techniques is an efficient

approach to detecting image forgery attacks.

iii
 ACKNOWLEDGMENTS

I want to express my gratitude to my advisor, Dr. Razvan Andonie, for his

guidance, support, and participation in the success of this thesis through all the steps

involved. I also want to thank my committee members: Dr. Szilard Vajda and Dr. Boris

Kovalerchuk for generously giving their time, support, advice, and goodwill during this

document’s review.

iv
 TABLE OF CONTENTS

Chapter Page

I INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

II THEORETICAL BACKGROUND . . . . . . . . . . . . . . . . . . . . . 5

Convolutional Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . 5

Support Vector Machines . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
Random Forests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
Image Pre-Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
Pre-Trained ImageNet Models and Transfer Learning . . . . . . . . . . . . . 11

III ERROR LEVEL ANALYSIS ALGORITHM . . . . . . . . . . . . . . . . 15

Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
The Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
Algorithm Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
Further Discussion on the Error Level Analysis Algorithm . . . . . . . . . . 18
Quality Dependence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

IV EXPERIMENTS WITH ERROR LEVEL ANALYSIS AND RESULTS . . 20

The CASIA Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

System Specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
Convolutional Neural Networks with ELA . . . . . . . . . . . . . . . . . . . 21
Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
Support Vector Machine with ELA . . . . . . . . . . . . . . . . . . . . . . 23
Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
Random Forests with ELA . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

v
Chapter Page

TABLE OF CONTENTS (CONTINUED)

V VGG16 EXPERIMENTS AND RESULTS . . . . . . . . . . . . . . . . . 25

RAISE Dataset and Reference Database . . . . . . . . . . . . . . . . . . . . 25

Image Pre-Processing Approach . . . . . . . . . . . . . . . . . . . . . . . . 25
VGG16 Pre-Trained Model . . . . . . . . . . . . . . . . . . . . . . . . . . 26

VI FUTURE WORK AND CONCLUSIONS . . . . . . . . . . . . . . . . . . 31

Future Work : Ground Truth Masks . . . . . . . . . . . . . . . . . . . . . . 31

Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

APPENDICES

A– CASIA DATASET . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

B– RAISE DATASET AND REFERENCE DATABASE . . . . . . . . . . . 41

C– RGB TABLE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

vi
 LIST OF FIGURES

Figure Page

1 Error Level Analysis model architecture . . . . . . . . . . . . . . . . . . . 7

2 Model architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

3 VGG16 pre-trained model architecture . . . . . . . . . . . . . . . . . . . 13

4 Error level analysis algorithm: Python implementation . . . . . . . . . . . 16

5 Pristine image with corresponding error level analysis . . . . . . . . . . . 17

6 Manipulated image with corresponding error level analysis . . . . . . . . . 17

7 The same image saved with different qualities . . . . . . . . . . . . . . . . 19

8 Testing with various numbers of epochs and batches . . . . . . . . . . . . 22

9 Support vector machine results . . . . . . . . . . . . . . . . . . . . . . . . 23

10 Random forest results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

11 Command to install CGvsPhoto package . . . . . . . . . . . . . . . . . . 25

12 Patch accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

13 Patch validation accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

14 Validation accuracy is 94.5% . . . . . . . . . . . . . . . . . . . . . . . . 27

15 Whole image training and validation accuracy . . . . . . . . . . . . . . . . 28

16 Whole image training and validation loss . . . . . . . . . . . . . . . . . . 28

17 Generating random spliced images . . . . . . . . . . . . . . . . . . . . . . 29

18 Detecting spliced area . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

19 Benchmark 1: personal image with spliced region . . . . . . . . . . . . . . 30

20 Benchmark 2: spliced personal image with low brightness . . . . . . . . . 30

21 Manipulated image with its ground truth mask . . . . . . . . . . . . . . . 31

22 Pristine image . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

vii
Figure Page

23 Copy-move image. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

24 Image splicing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

25 An image from the RAISE dataset . . . . . . . . . . . . . . . . . . . . . . 41

26 An image from the Reference Database . . . . . . . . . . . . . . . . . . . 42

27 RGB table . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

viii
 CHAPTER I

INTRODUCTION

With the advent of digital cameras and other smart devices, it has become easy for

anyone to manipulate an image. Some manipulations are not harmful, such as changing

the brightness of an image or converting it to black and white. On the other hand, some

manipulations can cause harm to others and defame them, especially politicians and

celebrities.

Image forgery is the process of manipulating a digital image to hide valuable or

essential content or to force the viewer to believe an idea [1]. It has been defined as the

process of manipulating an original digital image to either conceal its original identity

or create an entirely different image than what was originally intended by the user of the

digital platform [2]. Forged images can cause disappointment and emotional distress and

affect public sentiment and behavior [3]. Images can transmit much more information

than text. People tend to believe what they can see, and this affects their judgment, which

leads to a series of unwanted responses. Because fabrications have become widespread,

the urgency to detect forgeries has significantly increased. The copymove approach is

one of the most widely used forgery techniques. It copies a part of the image and pastes

it onto another part of the image. The technique itself is not harmful, but it can lead to

critical situations if someone uses it with malicious intent.

Image forgery is done mainly for malicious reasons. It serves to distort information,

spread immorality and fake news, obtain money fraudulently from an unsuspecting

audience, ruin the reputation of a popular celebrity or any other public figure, and

spread adverse political influence among the users of a digital platform. Therefore, clear

authentication of images and videos uploaded to digital media platforms, before they are

1
used in any way, makes it more difficult for digital information users to share information

[4]. Image forgery is often used by malicious people to ruin the reputation of public

figures. Image forgery, especially through Photoshop, can be used to display unethical

behavior in public figures. It is also sometimes an attempt to influence consumers of the

goods produced or the services offered by these public figures to shift to other markets

[5]. This forgery could also be used for political reasons against opponent politicians

or their agents, spreading images that portray their immoral side. This aims to convey a

message to the public regarding the lack of integrity of the subject. Image forgery often

leads to emotional problems for those whose images are released to public websites in

disregard for their privacy. There have been reports of suicide due to the leaking of private

images to public digital platforms, after which the victims undergo significant rebuke.

These deaths negatively affect society.

Image forgery is also sometimes used to cheat victims of their money in

increasingly common fraud schemes. The forged images are uploaded with embedded

text, purportedly from the owner of the original image, with instructions that end in

innocent people losing money. This is also done with images portraying people who are

in dire need of help, with intentions of fraudulently acquiring money from unsuspecting

members of the public. Society then ceases helping even those who are in genuine need of

help because of the fear of being swindled.

For all of these reasons, it is vital to develop methods of detecting whether an image

is forged and to locate the region of manipulation.

Related Work

Scientific studies on image forgery have provided various approaches to detecting

whether an image is manipulated. In [6], Bunk et al. discuss how resampling is a vital

2
signature of manipulated images. They proposed a method of detecting and locating

the area of manipulation on an image using resampling detectors with deep learning

classifiers. In [7], Abdalla et al. offered an approach to classifying whether an image is

forged that involved transfer learning. Bappy et al. [8] applied a long-short term memory

network with encoderdecoder architecture to detect and localize the area of manipulation.

Wang et al. [9] employed a feature pyramid network based on ResNet with Mask R-CNN

to identify and locate manipulated regions.

In [10], Rahmouni et al. classified an image by dividing it into 100 x 100

patches and passing these patches into a patch classifier (image pre-processing). They

subsequently trained the resulting complete images with a convolutional neural network

(CNN). In [11], Kaur and Manro pre-processed the images first. They then altered

the images to grayscale space and performed Gaussian pyramid decomposition from

time to time. Afterward, they detected image forgery using the block-based approach.

Amit Doegar [12] proposed a method involving AlexNets with support vector machines

(SVMs) to classify whether an image was forged without specifying the exact area of

manipulation. Zhou et al. [13] employed a pre-trained model, VGG16, with a steganalysis

rich model and CNNs.

In [14], Gupta et al. examined several block-matching algorithms, such as exact

match and robust match, and compared their performances. In [15], Shivakumar and

Baboo proposed an approach using the speeded-up robust features algorithm in parallel

with the K-dimensional tree algorithm to identify the manipulated region. Salloum

et al. [16] proposed using fully convolutional networks. In the beginning, they tried

using single-task fully convolutional networks. However, they noticed that multi-task

fully convolutional networks obtained better results than single-task fully convolutional

networks. The Pham et al. [17] segmented the manipulated images into spliced areas

3
and background areas in the manipulation-detection stage before the image-redemption

stage to improve the accuracy of the redemption. They suggested a hybrid approach that

could easily retrieve images using Zernike moment features and features found by a scale-

invariant feature transform.

Contributions

At the beginning, I began working on the error level analysis (ELA) algorithm

with various machine learning classifiers. A detailed discussion of the algorithm and the

classifiers follow in Chapter II, Chapter III, and Chapter IV. I continued researching this

area until I realized that the ELA algorithm is not the best way to detect image forgery.

I discovered that image pre-processing techniques, discussed in Chapter II and Chapter

V, obtain more accurate and promising results. I used the same pre-processing technique

presented in [10]. The authors of [10] divided each image into 100 x 100 patches and

passed them to a patch classifier that classifies whether the patch belongs to a raw image

or a computer graphics image. They passed the resulting images to a CNN to predict

the result. I used the same patch classifier technique, but instead of passing the resulting

images into a standard CNN, I passed them to a VGG16 pre-trained model. The authors

of [10] achieved high accuracy, 93.4%, with their model, but I obtained a higher accuracy,

94.5%, using VGG16.

4
 CHAPTER II

THEORETICAL BACKGROUND

Convolutional Neural Networks

Convolutional Neural Networks (CNNs or ConvNets) are a type of neural network

that are used effectively in image recognition classification and applications. Specifically,

these neural networks are effective in facial, traffic-sign, and object identification. They

also help power vision in robots and remote-controlled cars (self-driving) [18]. They

evolved from the LeNet architecture, which was the initial CNN that was useful in the

development of deep learning [19]. There are four operations that form the foundation of

every CNN:

1. An image composed of a matrix of pixel values

From computer graphics concepts, every image can be represented as pixel value

matrices. Certain components of an image are referred to as channels. In digital

camera images, three channels are present: red, green, and blue [20]. These three

channels are stacked in layers in the form of 2-dimensional matrices, and each

channel has a pixel value that is in the range of 0 to 255. Grayscale images are

distinguished by the presence of only one channel.

2. The convolution step

The convolution step entails extracting features from the input image. This

operation maintains the patterns between the pixels with the help of small squares

of input data that learn the features of the image. The operation involves sliding one

channel matrix, such as orange, by one pixel over the original image, which could

be green. This sliding step is referred to as a stride. Element-wise multiplication is

5
 performed for every position. Finally, these products are added to generate the final

integer that represents a single element of the desired output matrix, such as pink.

The operations in this block involve a filter and a convolved feature, which interact

on the input image to detect features from it.

3. Non-linearity (rectified linear unity [ReLU])

After every convolution operation, the ReLu involves a non-linear operation. The

ReLu operates on every pixel and replaces all negative pixel values in the feature

map with zero. The ReLu aims to introduce non-linearity to the CNN, as almost

every datum learned in the CNN has a linear property.

4. The pooling step

The pooling step retains the most significant information of the feature map while

reducing the dimensionality of every feature through spatial pooling [21]. The

pooling step involves pooling of different types such as average, sum, and max.

For example, if the operation involves max pooling, then the spatial neighborhood

must be defined, and subsequently, the largest element from the rectified feature

map within the window is selected [21]. The operation uses the average instead of

the largest element for average pooling and the sum for sum pooling. Therefore,

pooling, convolution, and ReLU are the foundation blocks for the effective

implementation of CNN.

Figure 1 shows our model architecture of the ELA algorithm as a pre-processing step and

the passing of the resulting images into a CNN. The algorithm is discussed in Chapter III.

6
 FIGURE 1: Error Level Analysis model architecture

Support Vector Machines

The Support Vector Machine (SVM) [22] is a model used in classification and

regression. It can solve linear and non-linear problems and performs well on various

practical challenges. The SVM algorithm generates a hyperplane that divides the data

into categories [23]. It is best applied in regression and classification problems, and it

produces the highest accuracy while using less computational power [24]. This algorithm

can be applied in classification, where the hyperplane in an N-dimensional space classifies

distinct data points. The SVM is classified as a supervised machine learning model. It

categorizes sets of training data into one or two other categories, and then a training

algorithm model is built to assign the categories to their respective groups. The SVM

is a non-probabilistic binary linear classifier that employs methods such as Platts scaling.

When working with textual analysis classification tasks, the SVM process involves

refining training data while employing other forms such as Naive Bayes algorithms.

A confidence score is generated for each recognized text or digit. When confidence is

achieved in the dataset, the SVM continues the classification by applying a classification

algorithm that is suitable when in situations with limited data. The algorithm involves

separating two classes of data points with various choices of hyperplane. The SVM

focuses on finding the plane with the maximum margin that represents the distance

7
between two data points in both classes. Classifying future data points becomes effective

with reinforcement from the maximization of the margin distance.

In SVMs, hyperplanes represent decision boundaries that are essential for data

point classification. The allocation of a data point to a certain class is based on the

side of the hyperplane that the point falls on. There are various features that form the

basis of the dimension of the hyperplane. Data points at the hyperplane determine the

orientation of the hyperplane to define support vectors. The margin of the classifier is

maximized through the support vectors. Support vector machines use the kernel trick to

perform linear classification while implicitly mapping inputs into feature spaces of high

dimension. The main goal of the SVM is to help classify data in most of the statistical

problems presented to machine learning experts. Understanding the correct position of

data points on the hyperplane makes it possible to apply the SVM effectively.

Support vector machines provide solutions to real problems in a wide range of

applications. A main application is text and hypertext categorization, which reduces the

requirement for labeled training in transductive and inductive settings. The classification

of images is another major area employing SVMs. The SVM is believed to achieve the

highest search accuracy compared to traditional query refinement techniques [25].

Random Forests

A random forest (RF) [26] is an ensemble algorithm, meaning that a decision

is made using the results from various models. In most cases, the outcome from an

ensemble model is better than that of any individual model [27]. Several decision trees

are generated by RFs, and the decision is determined based on the outcome of all the

decision trees [28]. An RF is a learning algorithm that randomly collects decision trees.

Each decision tree consists of several decisions, and a combination of them forms the

8
RF [29]. An RF integrates a collection of decision models from individual decision

trees in the forest to improve the accuracy of the results. This prevents relying on a

single learning model from a single decision tree in the RF. The merging of individual

decision trees, each with its own set of algorithms constituting the RF, therefore creates a

classification algorithm. This classification algorithm is independent from the algorithms

of the individual decision trees. This forms a basis of prediction that is more accurate than

that prediction that could have been made by a single decision tree or by a combination of

independent decision trees.

Decision trees are the building blocks of an RF, and the individual trees are used

to differentiate various events based on their most unique aspects [30]. An RF consists

of many trees that make the work more straightforward when complex sets of data are

involved. They work on the principle that many uncorrelated decision trees, if made

to operate as a group, will yield clearer and more accurate results than any individual

decision tree. This is possible because the decision trees protect each other from the

independent errors they make [31]. For an RF to work effectively, the decisions made

by the individual decision trees should have little or no correlation with each other. There

should also be unique signs in the differentiating features so that the models perform

better than a random guess.

Random forests utilize the idea of bagging, a process that allows decision trees to

sample from the dataset while making replacements, which result in different trees [32].

This is possible because individual decision trees are highly sensitive to the data that they

are trained on. Bagging therefore produces results where the individual trees not only

train on different sets of data but can also use different characteristics to make informed

decisions.

9
 Image Pre-Processing

Image pre-processing is the process of improving image data by performing various

operations on images and suppressing unwanted distortions in them. It can also be used

to enhance certain unique features in an image that are crucial to further processing.

Image processing may be a basic task, such as resizing [33]. For example, to feed an

image dataset into a deep learning model, all images must be of the same size. Other

pre-processing tasks include geometric and color conversion, or the transition of color to

grayscale; standardization; and data augmentation [34].

In this thesis, I used the pre-processing technique presented in [10]. The pre-

processing step takes every image in the dataset and divides it into 100 x 100 patches.

Subsequently, the patches are passed into a CNN classifier that classifies whether the

given patch belongs to a raw image (green) or a computer graphics image (red). After

the classification of patches, the complete images are generated again from the classified

patches. The authors of [10] passed the resulting images into another CNN to classify

whether an image is spliced. I used the VGG16 pre-trained model instead to accomplish

better accuracy than a standard CNN. Figure 2 shows our model architecture. The pre-

processing subfigure is from [10], and the VGG16 subfigure is from [35]. The datasets

and experiments are discussed in Chapter V.

10
 FIGURE 2: Model architecture

Pre-Trained ImageNet Models and Transfer Learning

A pre-trained ImageNet model is a model that has been trained on a significantly

large dataset to solve a problem that is similar to the problem I want the model to solve

[36]. I used a model pre-trained for a certain task on the ImageNet dataset. The initial

training of the model could have been done on a similar or very different domain, but the

ability to solve problems remains useful [37]. Studies on modern computer vision have

revealed that models that perform better on ImageNet usually perform better on other

vision tasks as well. It is common practice to use imported models, such as MobileNets

or VGG, due to the relatively high costs involved in training these models from scratch.

The task of importing, usually referred to as transfer learning, is not only effective but

also cost friendly to any profit-making institution. Transfer learning is also common

because pre-training a model requires a relatively large dataset for the model to extract

11
the characteristics required for the given task [19]. For instance, ImageNet contains over

one million images in 1000 categories [38]. A lack of data makes it difficult to train a

model from scratch and makes it necessary to import a pre-trained model. Another reason

for importing a pre-trained model instead of training one from scratch is the time required

to train a model from scratch, depending on the experience of the trainer. This is because

one must do many calculations and experiments before discovering a CNN architecture.

Pre-trained models are also commonly used because training a model from scratch

requires specific computational resources that might not be available. This also makes

it necessary to import a pre-trained model. A pre-trained model that is imported is usually

more efficient than a model that can be trained from scratch [39]. Pre-trained models are

more accurate in most cases because they have been trained on a large number of classes,

such as the 1000 classes in ImageNet. Employing a pre-trained model enhances their

suitability to work on a wider range of issues compared to a model that is trained from

scratch. Importing a pre-trained model is also advantageous because the most complex

work of optimizing the parameters has usually been completed; what remains is only fine-

tuning the model, a process that involves adjusting the hyperparameters to improve the

pre-trained model. Another advantage of the pre-trained model is that it uses fewer steps

before the convergence of the output [39]. This is because for a classification task, the

features to be extracted will be similar, and it thus requires less time. Prior to choosing

to import a pre-trained model, thorough research must be conducted on the problem in

question, and the keywords should be determined based on the type of dataset to be used.

This is because, depending on the complexity of the dataset, some models usually work

better than others. VGG16 is a CNN model that is used in large-scale image recognition.

It provides high accuracy testing using ImageNet, which consists of 100 classes of 14

12
million images. The model comprises 16 layers with weights, indicated by the value 16 in

VGG16. Figure 3 shows sample VGG16 architecture from [40].

FIGURE 3: VGG16 pre-trained model architecture

The input to the first convolutional layer of the VGG16 model is an RGB image

of fixed-size, 224 x 224. The image is passed through a stack of convolutional layers

that represent the use of a receptive field of size 3x3, the smallest size that can represent

bottom, up, left, right, and center. The VGG16 also makes use of 1 x 1 convolution filters

in several of its configurations. This configuration is the transformation of linear input

channels, which is followed by non-linearity. In this model, the convolutional stride

has a fixed value of 1 pixel. Therefore, any spatial padding of the convolutional layer

input must be preserved during convolution. A stack of convolutional layers forms the

foundation for fully connected layers. However, the stack is different from the fully

connected layers in architecture and depth. Each of the top two stacks comprises 4096

channels, and the third layer can perform a 1000-way ImageNet large-scale visual

recognition challenge classification, giving it 1000 channels. The soft max layer forms

the final layer that contributes to the configuration of the fully connected layers in the

networks [41]. In the presence of hidden layers in pre-trained models, the ReLU non-
13
linearity is the basic option that is applied. One of the challenges facing this model is

the need for increased memory due to a high consumption of space. The VGG16 is a

responsible model that assists machine learning experts in applying pre-trained networks

to improve the level of learning [42].

14
 CHAPTER III

ERROR LEVEL ANALYSIS ALGORITHM

Error Level Analysis (ELA) is a tool for exposing fabricated regions in JPEG

images [41]. It can help recognize manipulations of compressed (JPEG) images

by detecting the noise distribution present after resaving the image at a particular

compression rate.

Overview

Neal Krawetz invented the concept of Error Level Analysis for images when he

noticed how errors spread when a JPEG image is saved [43][44]. When cutting out a

section of an image and pasting it into another image, the ELA for the pasted section

often detects a more significant error, which means it is brighter than the rest of the

image. There are several implementations of this algorithm, but they all follow the same

steps.

The Algorithm

1. Compress the input image with a given compression rate and save it as a new

image.

2. Calculate the pixel-by-pixel difference between the original image and the new

image.

3. Store the difference in variable elaImg.

4. Compute the minimum and maximum pixel values for each band in the image and

store them in variable extrema.

15
 5. Compute the maximum pixel in extrema and store it in variable max.

6. Calculate the new scale by dividing 255 by the max.

7. Enhance the brightness of elaImg based on the resulting scale, then save and return

the resulting ELA image.

Figure 4 illustrates our Python implementation of this algorithm.

import os
from PIL import Image, ImageChops, ImageEnhance

def ToEla(path, quality):

filename = path
resavedImageName =filename.split(’.’)[0]+’.resaved.jpg’
EalImageName = filename.split(’.’)[0] + ’.ela.png’

img = Image.open(filename).convert(’RGB’)
img.save(resavedImageName, ’JPEG’, quality=quality)
imgResaved = Image.open(resavedImageName)

elaImg = ImageChops.difference(img, imgResaved)

extrema = elaImg.getextrema()
diff = max([ex[1] for ex in extrema])
if diff == 0:
diff = 1
scale = 255.0 / diff

elaImg = ImageEnhance.Brightness(elaImg).enhance(scale)
os.remove(resavedImageName)
return elaImg

FIGURE 4: Error level analysis algorithm: Python implementation

16
 Algorithm Output

Figure 5 and Figure 6 show pristine and manipulated images with their

corresponding error level analysis.

(a) Pristine image of a whale (b) ELA of pristine image

FIGURE 5: Pristine image with corresponding error level analysis

(a) Image of hybrid creature (b) ELA of image of hybrid creature

FIGURE 6: Manipulated image with corresponding error level analysis

17
 I loaded Figure 5a into the ELA algorithm, and the output is given in Figure 5b.

Figure 5b shows a nearly entirely black box, indicating that there is no noise in the image,

which means that the image is real and not manipulated.

I calculated the error level of Figure 6a , as shown in Figure6b. The noise in the

image is clear and indicates that the image of the hybrid animal is manipulated.

Further Discussion on the Error Level Analysis Algorithm

A variable quality level is used control the amount of compression of a JPEG

image. The amount of JPEG compression is typically measured as a percentage of the

quality level. An image at 100% quality has (almost) no loss, and a 1%-quality image

is a very low-quality image. In general, quality levels of 90% or higher are considered

high quality, 8090% is medium quality, and 7080% is low quality. Anything below 70%

is typically very low quality [45]. Low-quality images can reduce the ability of analysis

algorithms to detect modifications. The ELA algorithm works by resaving an image at a

known quality level, such as 75%, and during Step 3 in the algorithm, it then identifies the

amount of error introduced [45].

Quality Dependence

Saving an image with different quality levels affects the ELA algorithm, and this

leads to a distinct number of bright pixels. The lower the quality, the higher the number of

bright pixels. An image can be modified and saved in a quality lower than the quality

used in the ELA algorithm, and this makes it difficult to detect whether the image is

manipulated. Figure 7 shows an example of a pristine image that is saved with different

qualities. Subfigures (a) and (d) show the outcome of saving an image with 90% quality.

Subfigures (b) and (e) show the outcome of saving an image with 70% quality. Subfigures

18
(c) and (f) show the outcome of saving an image with 48% quality. From Figure 7, I

have concluded that the lower the quality, the greater the noise. This means that the ELA

algorithm is not always accurate.

(a) ELA of 90% quality (b) Result for 90% quality

(c) ELA of 70% quality (d) Result for 70% quality

(e) ELA of 48% quality (f) Result for 48% quality

FIGURE 7: The same image saved with different qualities

19
 CHAPTER IV

EXPERIMENTS WITH ERROR LEVEL ANALYSIS AND RESULTS

The CASIA Dataset

The CASIA dataset [46] contains three categories that make it an appropriate

dataset for this research:

1. Pristine images: Unspoiled images in their original form. Shown in Appendix A.

2. Copymove images: The manipulated region has been copied from the same image

and pasted on another area of the same image. Shown in Appendix A.

3. Spliced images: The manipulated region has been copied from a different image

and pasted on this image. Shown in Appendix A.

System Specifications

The specifications of the computer I used to conduct these experiments are the

following:

– Operating system:Ubuntu 18.04.4 LTS.

TM
– CPU: Intel R Core i7.

– RAM: 16 GB.

– GPU: NVIDIA GeForce 2060.

– Driver: NVIDIA Driver 430.

– External hard drive: 6 TB external hard drive.

20
 I tested the ELA algorithm on the CASIA dataset with three different classifiers:

CNNs [47], SVMs, and RFs. The following sections contain explanations of and results

for each classifier with the ELA algorithm.

Convolutional Neural Networks with ELA

The image paths were passed to a function that converts the images into their ELA

form, as discussed in Section 3.2. I then split the dataset into training and testing sets

using the train test split method from sklearn. I subsequently created a CNN

with three convolutional layers because it achieved better results than two layers. With

four layers, the model became overfitted. The results are discussed in Section 4.3. The

primary packages used include pandas to read images paths, matplotlib.pyplot

to plot performance curves, sklearn and keras to create the neural network.

The Image, ImageChops and ImageEnhance modules were used in the ELA

algorithm. Figure 8 shows validation and training accuracy for various numbers of epochs

and batches.

21
 Results

(a) 60 epochs batch size=50. (b) 60 epochs batch size=100.

(c) 100 epochs batch size=50. (d) 100 epochs batch size=100.

FIGURE 8: Testing with various numbers of epochs and batches

With the CNN, the model achieved a 79% accuracy. It had reached 80% when

I initially ran it on PopOs 19.04. However, I were required to downgrade the system

to Ubuntu 18.04 Bionic Beaver, since some Python packages and libraries are not yet

supported by PopOs. Figure 8 shows various performance measure curves: (a) shows

the accuracy of training with 60 epochs and a batch size of 50, (b) shows the accuracy of

training with 60 epochs and a batch size of 100, (c) shows the accuracy of training with

100 epochs and a batch size of 50, and (d) shows the accuracy of training with 100 epochs

and a batch size of 100. Of the training dataset, 80% was from the CASIA dataset. The

remaining 20% was for validation. Figure 8 shows that training accuracy was high. With
22
the validation data used on the model to evaluate its performance, the accuracy dropped to

72%.

Support Vector Machine with ELA

I converted all the images to their ELA format. I then passed the resulting two

folders (pristine image folder and manipulated image folder) to the SVM [48] with

rbf kernel. The results are shown in Section 4.4. The primary packages used include

skimage to open images and sklearn and keras to use the SVM classifier.

Results

The SVM achieved 72% accuracy, as shown in Figure 9 :

precision recall f1-score support

0 0.70 0.81 0.75 184
1 0.73 0.61 0.66 160
accuracy 0.72 344
macro avg 0.72 0.71 0.71 344
weighted avg 0.72 0.72 0.71 344

FIGURE 9: Support vector machine results

Random Forests with ELA

For feature extraction, I chose it based on the brightness of the pixels. I began

counting bright pixels on the ELA form. I estimated 150 pixels after consulting the

RGB table shown in Appendix C. According to the table, three zeros represent the color

black. The combination (255,255,255) represents white, the brightest shade. The second

row of the RGB table shows the combination (127,127,127) that represents gray. In my
23
opinion, the shade of gray provided by this combination was too dark. I therefore chose

the combination (150,150,150) as a boundary. Pixels brighter than (150,150,150) were

considered to be noise in the ELA images. The brighter the pixels in an ELA image,

the greater the noise. Figure 6b shows the ELA for a fake image. It contains 634 bright

pixels. I created the condition that if there are more than 300 bright pixels on an ELA

image, then the image is fake. I estimated 300 because some pristine images may have

some noise, as discussed in Section 3.4. The results are shown in Section 4.5. The

primary packages used include pandas to read images paths and sklearn to create

the RF.

Results

The RFs in this study achieved 91% accuracy, as shown in Figure 10 :

precision recall f1-score support

0 0.86 0.96 0.90 94
1 0.93 0.79 0.85 71
accuracy 0.91 165
macro avg 0.90 0.87 0.91 165
weighted avg 0.89 0.91 0.91 165

FIGURE 10: Random forest results

24
 CHAPTER V

VGG16 EXPERIMENTS AND RESULTS

RAISE Dataset and Reference Database

RAISE is a demanding real-world image data collection, developed primarily to test

digital forgery detection algorithms. It consists of approximately 8000 high-resolution

RAW images, which are uncompressed and never processed [49]. An image from this

dataset is shown in Appendix B. Reference Database is a free set of tagged screenshots

taken from games. It is a computer graphics database [50]. An image from this dataset is

shown in Appendix B. I conducted our experiments on 1800 randomly chosen images to

compare our results to those from [10]. The authors of [10] used CNNs to detect forgery

in an image. I used the same approach for image pre-processing but with a pre-trained

ImageNet model, VGG16, instead of a standard CNN.

Image Pre-Processing Approach

The authors of [10] made it possible to use their pre-processing approach by

providing the CGvsPhoto Python package.

pip3 install CGvsPhoto

FIGURE 11: Command to install CGvsPhoto package

As mentioned in Chapter II, the pre-processing function takes every image in the

dataset and divides it into 100 x 100 patches. The patches are then passed into a CNN

classifier that classifies whether the given patch belongs to a raw image (green) or a

25
computer graphics image (red). After classifying 100 patches, the pre-processing function

displays the accuracy of the classification. The classifier saves weights in a .ckpt file

after classifying 500 patches. After the classification of patches, the complete images are

regenerated from the classified patches. The authors of [10] passed the resulting images

into another CNN to obtain a final result, and they achieved an accuracy of 93.4%. I used

the vGG16 pre-trained model to accomplish higher accuracy than using a standard CNN.

For the pre-processing step, the authors of [10] achieved an accuracy of 84.4%.

VGG16 Pre-Trained Model

The authors of [10] achieved 84.4% accuracy on the pre-processing step. I obtained

87% on patch level. This higher accuracy was achieved because our weights were

optimized. Figure 12 shows the obtained patch accuracy.

FIGURE 12: Patch accuracy

Figure 13 shows the curve of validation accuracy.

26
 FIGURE 13: Patch validation accuracy

The authors of [10] passed the resulting complete images to another CNN and

achieved 93.4% accuracy by training 1800 images. I obtained 94.5% accuracy by using a

VGG16 pre-trained model. Figure 14 depicts our results.

FIGURE 14: Validation accuracy is 94.5%

Figure 15 and Figure 16 show the curves of training and validation.

27
FIGURE 15: Whole image training and validation accuracy

FIGURE 16: Whole image training and validation loss

28
 I wrote a program that generates randomly spliced images, illustrated in Figure17.

from PIL import Image, ImageDraw, ImageFilter

import os, random

export="/lubna/CWU/Thesis/generatingFakeImages/SpHuge/"
realPath="/lubna/CWU/Thesis/generateFakeImages/realimSP/"
realfiles=os.listdir(realPath)
fakePath="/lubna/CWU/Thesis/generatingFake/fakeimgSP/"
fakefiles=os.listdir(fakePath)

for i in range(1439):
left = 155
top = 65
right = 600
bottom = 600
x=random.randint(0,2300)
y=random.randint(0,2300)
a=random.choice(realfiles)
b=random.choice(fakefiles)
im1 = Image.open(realPath+a)
im2 = Image.open(fakePath+b)
im3=im2.crop((left, top, right, bottom))
backim = im1.copy()
backim.paste(im3, (x,y))
backim.save(export+str(i)+’.jpg’, quality=95)

FIGURE 17: Generating random spliced images

By using the image pre-processing function found in the library, I could locate the

spliced area in an image. I tested our model with a randomly generated spliced image,

and it was successful, as shown in Figure 18.

29
 FIGURE 18: Detecting spliced area

I also took a photograph from a personal camera and pasted an object onto it.

Figure 19 shows the result of detecting a spliced area in this photograph.

FIGURE 19: Benchmark 1: personal image with spliced region

I also attempted to input an image with lower brightness and a different object

location, as shown in Figure 20.

FIGURE 20: Benchmark 2: spliced personal image with low brightness

30
 CHAPTER VI

FUTURE WORK AND CONCLUSIONS

Future Work : Ground Truth Masks

Generating the ground truth mask of a manipulated image is more reliable than the

ELA algorithm, and it is not affected by the quality of the image. It also shows the exact

area of manipulation. Obtaining the ground truth mask requires knowledge of image pre-

processing and image segmentation techniques. Figure 21 shows a manipulated image

with its ground truth mask.

(a) Manipulated image.

(b) Ground truth mask.

FIGURE 21: Manipulated image with its ground truth mask

31
 Conclusions

Image forgery involves distorting images, sometimes images of people, for

malicious reasons. This involves a genuine image that had been displayed on a public

website or a digital communication platform and is edited into an entirely different image.

The new image will likely be immoral in nature or targeted to spread negative publicity.

The ELA algorithm shows whether an image is manipulated when the input images

quality is close to the quality used in the algorithm. If there is a large difference between

the quality of the image and the quality of the algorithm, then the result will always be

incorrect. Furthermore, the algorithm does not show the exact area of manipulation.

A pre-trained model is a model that has been trained on a certain task on the

ImageNet dataset. It is a model that has been trained to solve issues that might be similar

to the problem at hand. A pre-trained model is preferred in most cases to training a model

from scratch. The process of importing a pre-trained model is referred to as transfer

learning.

Other approaches do not depend on the quality of the images and show the exact

area of manipulation. The patch classification approach is not affected by the quality of

the image and achieves more accurate results. Commonly imported models such as VGG

and MobileNets have been trained on large sets of data and are therefore very efficient

on any given dataset. The time required to train a model from scratch, depending on

experience, is relatively high, which makes it necessary to consider using pre-trained

models.

32
 Bibliography

[1] M. Sridevi, C. Mala, and S. Sanyam, “Comparative study of image forgery

and copy-move techniques,” in Advances in Computer Science, Engineering &

Applications, D. C. Wyld, J. Zizka, and D. Nagamalai, Eds. Berlin, Heidelberg:

Springer Berlin Heidelberg, 2012, pp. 715–723.

[2] S. Walia and K. Saluja, “Digital image forgery detection: a systematic scrutiny,”

Australian Journal of Forensic Sciences, pp. 1–39, 03 2018.

[3] C. Shen, M. Kasra, W. Pan, G. A. Bassett, Y. Malloch, and J. F. OBrien, “Fake

images: The effects of source, intermediary, and digital media literacy on contextual

assessment of image credibility online,” New Media & Society, vol. 21, no. 2, pp.

438–463, 2019. [Online]. Available: https://doi.org/10.1177/1461444818799526

[4] C. N. Bharti and P. Tandel, “A survey of image forgery detection techniques,”

2016 International Conference on Wireless Communications, Signal Processing and

Networking (WiSPNET), pp. 877–881, 2016.

[5] C. Salge, “Is that social bot behaving unethically?” Communications of the ACM,

vol. 60, pp. 29–31, 08 2017.

[6] J. Bunk, J. H. Bappy, T. M. Mohammed, L. Nataraj, A. Flenner, B. S. Manjunath,

S. Chandrasekaran, A. K. Roy-Chowdhury, and L. Peterson, “Detection and

localization of image forgeries using resampling features and deep learning.” in CVPR

Workshops. IEEE Computer Society, 2017, pp. 1881–1889. [Online]. Available:

http://dblp.uni-trier.de/db/conf/cvpr/cvprw2017.html#BunkBMNFMCRP17

33
 [7] Y. Abdalla, M. Iqbal, and M. Shehata, “Image forgery detection based

on deep transfer learning,” European Journal of Electrical Engineering and

Computer Science, vol. 3, no. 5, Sep. 2019. [Online]. Available: https:

//ejece.org/index.php/ejece/article/view/125

[8] J. H. Bappy, C. Simons, L. Nataraj, B. S. Manjunath, and A. K. Roy-Chowdhury,

“Hybrid lstm and encoderdecoder architecture for detection of image forgeries,”

IEEE Transactions on Image Processing, vol. 28, no. 7, pp. 3286–3300, July 2019.

[9] X. Wang, H. Wang, S. Niu, and J. Zhang, “Detection and localization of image

forgeries using improved mask regional convolutional neural network.”

[10] N. Rahmouni, V. Nozick, J. Yamagishi, and I. Echizen, “Distinguishing computer

graphics from natural images using convolution neural networks,” in 2017 IEEE

Workshop on Information Forensics and Security (WIFS), 2017, pp. 1–6.

[11] G. Kaur and D. R. Manro, “A brief review : Copy-move forgery detection,” 2018.

[12] K. G. Amit Doegar, Maitreyee Dutta, “Cnn based image forgery detection using

pre-trained alexnet model,” Ph.D. dissertation, CHANDIGARH, India, 2020.

[13] P. Zhou, X. Han, V. I. Morariu, and L. S. Davis, “Learning rich features for image

manipulation detection,” in The IEEE Conference on Computer Vision and Pattern

Recognition (CVPR), June 2018.

[14] A. Gupta, N. Saxena, and S. Kumar, “Detecting copy move forgery in digital images,”

International Journal of Engineering Research and Applications, vol. 3, no. 1, pp.

94–97, 03 2013.

34
[15] B. Shivakumar and S. Baboo, “Detection of region duplication forgery in digital

images using surf,” International Journal of Computer Science Issues, vol. 8, no. 3,

pp. 199–205, 07 2011.

[16] R. Salloum, Y. Ren, and C.-C. J. Kuo, “Image splicing localization using a multi-task

fully convolutional network (mfcn),” J. Vis. Commun. Image Represent., vol. 51, pp.

201–209, 2018.

[17] N. T. Pham, J.-W. Lee, G.-R. Kwon, and C.-S. Park, “Hybrid image-retrieval method

for image-splicing validation,” Symmetry, vol. 11, no. 1, p. 83, 2019.

[18] A. L. Caterini and D. E. Chang, Deep Neural Networks in a Mathematical

Framework, 1st ed. Springer Publishing Company, Incorporated, 2018.

[19] K. Kang, “Comparison of face recognition and detection models: Using different

convolution neural networks,” Opt. Mem. Neural Netw., vol. 28, no. 2, p. 101108,

Apr. 2019. [Online]. Available: https://doi.org/10.3103/S1060992X19020036

[20] Y. Mao, Z. He, Z. Ma, X. Tang, and Z. Wang, “Efficient convolution neural networks

for object tracking using separable convolution and filter pruning,” IEEE Access,

vol. 7, pp. 106 466–106 474, 2019.

[21] K. Uchida, M. Tanaka, and M. Okutomi, “Coupled convolution layer for

convolutional neural network,” in 2016 23rd International Conference on Pattern

Recognition (ICPR), 2016, pp. 3548–3553.

[22] Wikipedia, “Support-vector networks,” pp. 273–297, 1995.

[23] R. Pupale, “Support Vector Machines(SVM) An Overview,” https://

towardsdatascience.com/, 2018.

35
[24] W. Land and J. Schaffer, The Support Vector Machine, 01 2020, pp. 45–76.

[25] X. Zhang, Support Vector Machines, 01 2017, pp. 1214–1220.

[26] Tin Kam Ho, “Random decision forests,” pp. 278–282 vol.1, 1995.

[27] Sklearn.ensemble.RandomForestClassifier, “sklearn Random forest

Classifier,” https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.

RandomForestClassifier.html, 2020.

[28] N. Horning, “Random forests: An algorithm for image classification and generation

of continuous fields data sets,” in Proceedings of the International Conference on

Geoinformatics for Spatial Infrastructure Development in Earth and Allied Sciences,

Osaka, Japan, vol. 911, 2010.

[29] R. Pandya and J. Pandya, “C5. 0 algorithm to improved decision tree with

feature selection and reduced error pruning,” International Journal of Computer

Applications, vol. 117, pp. 18–21, 05 2015.

[30] B. Gregorutti, B. Michel, and P. Saint-Pierre, “Correlation and variable importance

in random forests,” Statistics and Computing, vol. 27, 10 2013.

[31] Q. Zhang, Y. Yang, H. Ma, and Y. Wu, “Interpreting cnns via decision trees,” 06

2019, pp. 6254–6263.

[32] L. Ma, B. Sun, and Z. Li, “Bagging likelihood-based belief decision trees,” 07 2017.

[33] K. Pal and K. Sudeep, “Preprocessing for image classification by convolutional

neural networks,” 05 2016, pp. 1778–1781.

[34] M. Elgendy, Deep Learning for Vision Systems.

36
[35] GeeksforGeeks, “VGG-16 — CNN model,” shorturl.at/tvAKM, 2019.

[36] S. Hinterstoisser, V. Lepetit, P. Wohlhart, and K. Konolige, On Pre-trained Image

Features and Synthetic Images for Deep Learning: Munich, Germany, September

8-14, 2018, Proceedings, Part I, 01 2019, pp. 682–697.

[37] M. Simon, E. Rodner, and J. Denzler, “Imagenet pre-trained models with batch

normalization,” vol. 21, no. 1, pp. 2–4, 12 2016.

[38] M. Long, H. Zhu, J. Wang, and M. I. Jordan, “Deep transfer learning with joint

adaptation networks,” p. 22082217, 2017.

[39] K. He, R. Girshick, and P. Dollar, “Rethinking imagenet pre-training,” in 2019

IEEE/CVF International Conference on Computer Vision (ICCV), vol. 8, no. 2, 2019,

pp. 4917–4926.

[40] R. Thakur, “Step by step VGG16 implementation in Keras for beginners,” shorturl.at/

gjzL5, 2019.

[41] A. S. Pankaj Kumar Kandpal, Ashish Mehta, “Honey bee bearing pollen and

non-pollen image classification, vgg16 transfer learning method using different

optimizing functions,” International Journal of Innovative Technology and Exploring

Engineering (IJITEE), vol. 57, pp. 2–5, 12 2019.

[42] A. Omar, “Lung ct parenchyma segmentation using vgg-16 based segnet model,”

International Journal of Computer Applications, vol. 178, pp. 10–13, 08 2019.

[43] I. Steadman, “ ’Fake’ World Press Photo isn’t fake, is lesson in need for forensic

restraint,” https://www.wired.co.uk/article/photo-faking-controversy, 2012.

[44] S. Masters, “ Error Level Analysis,” shorturl.at/cfuBE, 2016.

37
[45] Fotoforensics, “Tutorial: Error Level Analysis,” https://fotoforensics.com/

tutorial-ela.php, 2012-2019.

[46] Y. Z. IEEE Dataport, “CASIA Dataset,” http://dx.doi.org/10.21227/c1h8-kf39, 2019.

[47] I. Zafar, G. Tzanidou, R. Burton, N. Patel, and L. Araujo, Hands-On Convolutional

Neural Networks with TensorFlow: Solve Computer Vision Problems with Modeling

in TensorFlow and Python. Packt Publishing, 2018.

[48] Sklearn, “Support Vector Machines,” https://scikit-learn.org/stable/modules/svm.

html, 2020.

[49] D.-T. Dang-Nguyen, C. Pasquini, V. Conotter, and G. Boato, “Raise: A

raw images dataset for digital image forensics,” in Proceedings of the 6th

ACM Multimedia Systems Conference, ser. MMSys 15. New York, NY, USA:

Association for Computing Machinery, 2015, p. 219224. [Online]. Available:

https://doi.org/10.1145/2713168.2713194

[50] L. Design, “Reference Database,” http://level-design.org/referencedb/, 2009-2020.

[51] T. Binary, “Convert Truecolor To Binary Online,” shorturl.at/quQX4, 2019.

38
 APPENDIX A

CASIA DATASET

FIGURE 22: Pristine image

In Figure 23, the top right flower has been cut, resized, and pasted onto the lower

left section of the image.

FIGURE 23: Copy-move image.

39
 In Figure 24, the two yellow flowers in the top left have been cut from Figure 24a

and pasted onto Figure 24b. The result is shown in Figure 24c.

(a) Source (b) Destination (c) Spliced image

FIGURE 24: Image splicing

40
 APPENDIX B

RAISE DATASET AND REFERENCE DATABASE

FIGURE 25: An image from the RAISE dataset

41
FIGURE 26: An image from the Reference Database

42
 APPENDIX C

RGB TABLE

Figure 27 shows the RGB table from [51]

FIGURE 27: RGB table

43