Abstract

Digital image modification or image forgery is easy to do today. The authenticity verification
of an image become important to protect the image integrity so that the image is not being
misused. Error Level Analysis (ELA) can be used to detect the modification in image by
lowering the quality of image and comparing the error level. The use of deep learning
approach is a state-of-theart in solving cases of image data classification. This study wants to
know the effect of adding ELA extraction process in the image forgery detection using deep
learning approach. The Convolutional Neural Network (CNN), which is a deep learning
method, is used as a method to do the image forgery detection. The impacts of applying
different ELA compression levels, such as 10, 50, and 90 percent, were also compared in this
study. According to the results, adopting the ELA feature increases validation accuracy by
about 2.7% and give the better test accuracy. However, the use of ELA will slow down the
processing time by about 5.6%.

Introduction:
Introduction Image data is extremely prone to manipulation in today's digital environment.
Image editing software is widely available today, and it may be used not only on desktop
computers and laptops, but also on handheld mobile devices. A deep generative model is
commonly used to create hyper-realistic face-swapping images and videos in some
applications [2]. People commonly exploit the outcomes of this image manipulation on social
media, in the commercial world, and even for criminal purposes. The use of image
manipulation for illegal purposes should be a major source of concern, since it can pose a
serious threat to society, government, and industry. As a result, the validity of the images in
the internet must be confirmed. So that, maintaining the integrity of digital photographs is
crucial [3]. In this situation, an image forgery detection method can be used to verify the
validity of digital images [4]. The Digital Image Forensics (DIF) is a field to detect the
authenticity of digital images, both in terms of the image content's integrity and the source
[5]. Active and passive modification detection techniques are two types of algorithms for
detecting picture forgery in DIF [6]. The method of passive forgery detection does not need
any prior knowledge of the image's content [7]. The active technique, on the other hand,
requires extracting watermarks and digital signatures embedded in images and then verifying
them [8]. As a result, any modification to the image can disrupt the embedded watermark and
digital signature, assisting in the detection of the image's validity. The copy-move (cloning)
modification method is the passive image forgery that has the largest impact on the original
image.

Related Work:

Various approaches have been proposed in the literature to deal with image forgery. The
majority of traditional techniques are based on particular artifacts left by image forgery,
whereas recently techniques based on CNNs and deep learning were introduced, which are
mentioned below. First, we will mention the various traditional techniques and then move on
to deep learning based techniques. In [14], the authors’ proposed error level analysis (ELA)
for the detection of forgery in an image. In [15], based on the lighting conditions of objects,
forgery in an image is detected. It tries to find the forgery based on the difference in the
lighting direction of the forged part and the genuine part of an image. In [16], various
traditional image forgery detection techniques have been evaluated. In [17], Habibi et al., use
the contourlet transform to retrieve the edge pixels for forgery detection. In [18], Dua et al.,
presented a JPEG compression-based method. The discrete DCT coefficients are assessed
independently for each block of an image partitioned into non-overlapping blocks of size 8 ×
8 pixels. The statistical features of AC components of block DCT coefficients alter when a
JPEG compressed image tampers. The SVM is used to classify authentic and forged images
using the retrieved feature vector. Ehret et al. in [19] introduced a technique that relies on
SIFT, which provides sparse keypoints with scale, rotation, and illumination invariant
descriptors for forgery detection. A method for fingerprint faking detection utilizing deep
Boltzmann machines (DBM) for image analysis of high-level characteristics is proposed in
[20]. Balsa et al. in [21] compared the DCT, Walsh–Hadamard transform (WHT), Haar
wavelet transform (DWT), and discrete Fourier transform (DFT) for analog image
transmission, changing compression and comparing quality. These can be used for image
forgery detection by exploring the image from different domains. Thanh et al. proposed a
hybrid approach for image splicing in [22], in which they try to retrieve the original images
that were utilized to construct the spliced image if a given image is proven to be the spliced
image. They present a hybrid image retrieval approach that uses Zernike moment and SIFT
features.

Error Level Analysis (ELA):

 Error Level Analysis is a technique used to detect image manipulation by restoring the image
at a certain quality level and calculating the ratio between compression levels [26]. In
general, this technique is performed on images that are in a lossy format (lossy compression).
The image type used in data mining is JPEG. In JPEG images, compression is performed
independently for every 8 × 8 pixel in the image. If an image is not manipulated, every 8 × 8
pixels in the image must have the same error rate [23]. Three different levels of image
compression were examined in this article, namely 10%, 50%, and 90% compression.

Methodology:

Convolution Layer:

 Convolutional neural networks apply a filter to an input to create a feature map that summarizes the
presence of detected features in the input.

Convolutional Neural Networks have the following layers:

o Convolutional
o ReLU Layer
o Pooling
o Fully Connected Layer

Step1: Convolution Layer:

 Convolutional neural networks apply a filter to an input to create a feature map that summarizes the
presence of detected features in the input.
Step2 : ReLU Layer

In this layer, we remove every negative value from the filtered images and replaces them with
zeros.It is happening to avoid the values from adding up to zero.

Rectified Linear unit(ReLU) transform functions only activates a node if the input is above a certain
quantity. While the data is below zero, the output is zero, but when the information rises above a
threshold. It has a linear relationship with the dependent variable.

Step3: Pooling Layer

In the layer, we shrink the image stack into a smaller size. Pooling is done after passing by
the activation layer. We do by implementing the following 4 steps:

o Pick a window size (often 2 or 3)

o Pick a stride (usually 2)
o Walk your Window across your filtered images
o From each Window, take the maximum value

Step4: Fully Connected Layer

The last layer in the network is fully connected, meaning that neurons of preceding layers are
connected to every neuron in subsequent layers.

This mimics high-level reasoning where all possible pathways from the input to output are
considered.

Then, take the shrunk image and put into the single list, so we have got after passing through two
layers of convolution relu and pooling and then converting it into a single file or a vector.

Hardware Requirements

Processor : Intel i3 3.30 GHz.

Hard Disk : 100 GB (min)

Ram : 8GB

Software Requirements:

Language : Python

Software : Anaconda

Editor : Visual Studio Code

Framwerok : Flask

Notebook : Jupyter