Broadcasting Forensics Using Machine Learning Approaches

International Journal of Trend in Scientific Research and Development (IJTSRD)
Volume 7 Issue 3, May-June 2023 Available Online: www.ijtsrd.com e-ISSN: 2456 – 6470
Broadcasting Forensics Using Machine Learning Approaches

Amit Kapoor1, Prof. Vinod Mahor2
1
M Tech Scholar, 2Assistant Professor,
1,2
Computer Science & Engineering, Millennium Institute of Technology & Science, Bhopal, Madhya Pradesh, India
ABSTRACT How to cite this paper: Amit Kapoor |

Broadcasting forensic is the practice of using scientific methods and Prof. Vinod Mahor "Broadcasting
techniques to analyse and authenticate Multimedia content. Over the Forensics Using Machine Learning
past decade, consumer-grade imaging sensors have become Approaches"
Published in
increasingly prevalent, generating vast quantities of images and
International Journal
videos that are used for various public and private communication
of Trend in
purposes. Such applications include publicity, advocacy, Scientific Research
disinformation, and deception, among others. This paper aims to and Development
develop tools that can extract knowledge from these visuals and (ijtsrd), ISSN: 2456- IJTSRD57545
comprehend their provenance. However, many images and videos 6470, Volume-7 |
undergo modification and manipulation before public release, which Issue-3, June 2023, pp.1034-1045, URL:
can misrepresent the facts and deceive viewers. To address this issue, www.ijtsrd.com/papers/ijtsrd57545.pdf
we propose a set of forensics and counter-forensic techniques that
can help establish the authenticity and integrity of Multimedia Copyright © 2023 by author (s) and
content. Additionally, we suggest ways to modify the content International Journal of Trend in
intentionally to mislead potential adversaries. Our proposed tools are Scientific Research and Development
Journal. This is an
evaluated using publicly available datasets and independently
Open Access article
organized challenges. Our results show that the forensics and distributed under the
counter-forensic techniques can accurately identify manipulated terms of the Creative Commons
content and can help restore the original image or video. Attribution License (CC BY 4.0)
Furthermore, in this paper demonstrate that the modified content can (http://creativecommons.org/licenses/by/4.0)
successfully deceive potential adversaries while remaining
undetected by state-of-the-art forensic methods.
KEYWORDS: Broadcasting Forensics, Multimedia Content Analysis,
Image and Video Manipulation Authenticity and Integrity, Forensics
and Counter-Forensic Techniques, ML
1. INTRODUCTION
The widespread use of smartphones has news" problem [1] and the recent rise in popularity of
fundamentally altered how we communicate with one machine learning-based techniques like generative
another. Unprecedented levels of digital content adversarial networks that may be abused by bad
creation, such as photographs and films, are now actors are also aggravating this scepticism.
possible because to these gadgets. The gold standard
The research community has been creating methods
for storing our memories or indicating our
to check and authenticate such content in order to
participation at social occasions is currently thought
reduce the spread of false information and assist law
to be digital information. Numerous pieces of
enforcement authorities with the completion of
software have been created in order to edit and
investigations in which digital assets may be
improve these digital assets because of the value
involved. These initiatives gave rise to the area of
placed on images and videos as well as their
digital image and video forensics [2]. The present
significant role in social networks and journalistic
state of digital image and video forensics includes the
Broadcasting. This software innovation has a cost,
two significant study topics of source image forensics
though. Image and video manipulation, whether done
and content-based forensics, which are the subject of
for aesthetic or malevolent reasons, is becoming a
this work. We specifically frame the issue of source
popular practise. The public is starting to reject taking
image forensics as a source camera identification
any image or video as digital evidence because of the
challenge, which we suggest resolving using CNN-
abundance of these faked Broadcasting. One of the
based techniques. Computer graphics forensics, on
first to make a suggestion of this pattern. The "fake
@ IJTSRD | Unique Paper ID – IJTSRD57545 | Volume – 7 | Issue – 3 | May-June 2023 Page 1034
International Journal of Trend in Scientific Research and Development @ www.ijtsrd.com eISSN: 2456-6470
which we specifically focus on the current Deepfakes of source image forensics, deep learning-based
video manipulation approach, are a subset of content- techniques have been effectively used, and they have
based forensics. We also research the production and demonstrated their efficacy in a number of contests
detection of hostile content images for CNN camera [4]. In order to better comprehend the outcomes
model identification techniques. The use of video produced by deep learning techniques, we provide a
metadata analysis for tampering detection is thorough analysis of those methods in the lines that
something else we look into [3]. follow, as well as their interrelationships, as seen in
Figure 1.1 Any reader of this work will be able to
The analysis of the picture's statistics and other
pattern recognition techniques were initially used to better understand the context in which we developed
our solutions as a result of this.
solve the issue of source image forensics. In the field
Figure 1.1: The Venn diagram demonstrating the relationship between deep learning and
representational learning.
Deep neural networks became practical in thanks to a number of innovations, which also gave rise to the area of
deep learning. Following deep learning-based techniques has consistently produced outstanding results on a
variety of tasks, including handwritten text recognition, understanding video scenes, and image classification.
Convolutional neural networks, recurrent neural networks, and generative adversarial networks are some of the
specialised methods in computer vision that have been developed to address these problems. Among these,
CNNs and RNNs have been demonstrated to be efficient when handling image- and video-related tasks, and as a
result, they have been used as the foundation for several digital forensics techniques [5–6].
Convolutional layers, pooling layers, and activation functions are the primary fundamental elements of CNNs.
These layers are often layered together to create the convolutional neural network's architecture. The four
important recent developments in the construction and training of CNNs are regularisation, structural
reformulation, and loss function. The most crucial of these for enhancing performance is structural reformulation
[7].
Figure 1.2: Deep CNN architecture taxonomy.

1.1. Description of the Camera Model
Identification of the camera model is essential for blindly determining the source and validity of a picture.
Modern methods make use of characteristics from the image acquisition pipeline that describe the traces that
various camera models (such as those left by proprietary demos icing techniques) left on pictures. A
convolutional neural network (CNNs) might be used for camera model recognition in light of the very accurate
performance attained by feature-based approaches as well as the advancements brought about by deep
architectures in machine learning. We especially look at (i) how well different network topologies can learn
discriminant characteristics from seen pictures, and (ii) how much training data is necessary to obtain a certain
level of accuracy [8].
2. LITERATURE SURVEY
The study and identification of unlicensed broadcast content, known as broadcasting forensics, is a challenge
that is getting harder to solve in the modern world. Due to its capacity to automatically analyse and discover
abnormalities in broadcast information, the usage of machine learning algorithms has grown in popularity as a
remedy for this issue. We will examine some of the most recent studies on the application of machine learning to
broadcasting forensics in this review of the literature.
He and Zhang released one of the first publications in this area in 2015, outlining a machine learning-based
method for locating unauthorised films on social Broadcasting sites. Based on a collection of data taken from the
video frames, the authors utilised a supervised learning technique called Support Vector Machine (SVM) to
categorise films as authorised or unauthorised. A lot of data is available on social Broadcasting, which is
growing essential for communication, and this platform might be useful for analysis. We use Twitter to tweet
about seemingly unrelated subjects and various emotions at specific times. In this article, sentimental analysis is
suggested as a way to learn about people's thoughts or feelings. Every tweet is divided into its suitable emotion.
Either a pleasant or negative feeling might be present. The two phases of the suggested technique are pre-
processing and categorization. After completing all essential pre-processing, the corpus is formed. For
classification, techniques including Logistic Regression, Linear SVC, Bernoulli NB, Decision Tree Classifier,
Voting Classifier, and KNN Classifier are employed. Data from Twitter for 2020 and 2021 has been collected
for testing. Linear SVC performs better in terms of accuracy on training data, whereas linear regression performs
better in terms of accuracy on testing data [7-8].
Chiang and associates proposed a comparable strategy for identifying unapproved audio material in radio
broadcasts in 2022. Based on their spectrogram representations, the authors employed a Convolutional Neural
Network (CNN) to categorise audio portions as authorised or unauthorised. We need to gather artefacts from
many systems, applications and databases, network and security devices, and other things as part of digital
forensics and incident response. The suggested technique obtained an accuracy of over 95% in detecting
unauthorised audio segments. Artefacts will assist us in comprehending contemporary and historical events that
took place in a certain system at a specific period. Event Logs and Registry Values, particularly USB, which
offers details on external Broadcasting linked to the computer, and Run Keys, which provides details on the
virus persistence mechanism, are two of these artefacts on Windows machines. It's conceivable for an attacker to
run a malicious power shell script that controls the system or to leave a back door power shell script that allows
access to the victim's system. In order to investigate if attackers are employing power shell, we are providing a
method to gather power shell events. The first step in identifying a danger or an attack in its entirety is to
examine security event logs. As a result, if a user receives a spam email, security events will produce a few
associated events, and we will also use the SIEM tool for additional analysis. Our technology also converts all
event logs to text files, making it simple for an investigator to review those files in a SIEM tool. All monitoring
events, such as the beginning and ending of processes, network connections, any modifications to file creation,
timestamps, etc., are recorded in the Sysmon event log. The sysmon event logs may be fetched by this Python-
based utility and saved in a text file. A researcher can submit the file to the SIEM tool for additional study. This
may provide the current circumstance a new security patch [9-10].
3. PROBLEM IDENTIFICATION
In this section we introduce the problem formulation with the notation used throughout the chapter. We then
provide the reader a brief overview about CNNs and their use in Multimedia Forensics.
3.1. Problem Formulation
Let us consider a color image I acquired with camera model l belonging to a set of known camera models L. In
this chapter, we consider the patch-based closed-set camera model attribution problem as presented in [6]. Given
an image I, this means:
Select a subset of K color patches Pk, k ∈ [1, K].
Obtain an estimate ˆlk = C(Pk) of the camera model associated with each patch through a camera attribution
function C.
Optionally obtain final camera model estimate ˆl through majority voting over ˆlk, k ∈ [1, K].
Our goal is to detect whether a patch Pk is a good candidate for camera model attribution estimation. To this
purpose, we propose a CNN architecture that learns a function G expressing the likelihood of a patch Pk to
provide correct camera model identification, i.e., gk = G(Pk). High values of gk indicate high probability of
patch Pk to provide correct camera information. Conversely, low gk values are attributed to patches Pk that
cannot be correctly classified. Pixel-wise likelihood is then represented by means of a reliability map M,
showing which portion of an image is a good candidate to estimate image camera model, as shown in Figure 3.1.
3.2. Convolutional Neural Networks in Multimedia Forensics
In this section, we present a brief overview of the foundations of convolutional neural networks (CNNs) that are
needed to follow the chapter. For a thorough review on CNNs, we refer the readers of this paper.
Deep learning and in particular CNNs have shown very good performance in several computer vision
applications such as visual object recognition, object detection and many other domains such as drug discovery
and genomics [11]. Inspired by how the human vision works, the layers of a convolutional network have neurons
arranged in three dimensions, so each layer has a width height, and depth. The neurons in a convolutional layer
are only connected to a small, local region of the preceding layer, so we avoid wasting resources as it is common
in fully-connected neurons. The nodes of the network are organized in multiple stacked layers, each performing
a simple operation on the input. The set of operations in a CNN typically comprises convolution, intensity
normalization, non-linear activation and thresholding, and local pooling. By minimizing a cost function at the
output of the last layer, the weights of the network are tuned so that they are able to capture patterns in the input
data and extract distinctive features. CNNs enable learning data-driven, highly representative, layered
hierarchical image features from sufficient training data.
Figure 3.1: A block schematic illustrating the suggested strategy. Image I'm divided into patches.
There has been a growing interest in using convolutional neural networks in the fields of image forensics and
steganalysis [12]. These papers mainly focus on architectural design of CNNs where a single CNN model is
trained and then tested in experiments. Data- driven models have recently proved valuable for other Multimedia
forensic applications as well, Moreover, initial exploratory solutions targeting camera model identification show
that it is possible to use CNNs to learn discriminant features directly from the observed known images, rather
than having to use hand-crafted features. As a matter of fact, the use of CNNs also makes it possible to capture
characteristic traces left by non- linear and hard to model operations present in the image acquisition pipeline of
capturing devices.
Results have indicated that learning from interBroadcastingte representation in CNNs instead of output
probabilities, and then jointly retraining the final architecture, leads to performance improvement [13].
Figure 3.2: Representation of the proposed CNN architecture M engaged in image patch processing
3.3. Estimating Patch Reliability Method
In this part, we go into further depth about how we estimate patch reliability and attribute cameras. The
following phases make up the proposed pipeline (see Figure 3.2):
A. Patches are separated from the image being analysed.
B. To calculate the likelihood of patch reliability, a CNN is utilised.
C. We estimate a camera model for each patch using the same CNN.
D. The camera is attributed to the entire image, and a dependability mask is created.
E. A thorough description of each step is provided below.
F. Patch Extraction
The suggested technique examines picture patches to function. The colour picture I is initially divided into a
collection of K patches Pk, k [0, K]. The resolution of each patch is 64 by 64 pixels. Depending on the
dimension, the patch extraction stride might range from 1 to 64.Mr. Sri Kalyan Yarlagadda, Prof. Fengqing
Maggie Zhu, Prof. Paolo Bestagini, and Prof. Stefano Tubaro collaborated on this project.
4. PROPOSED METHOD
The proposed counter-forensic method. Our method consists of an adversarial image generator module that can
be added to a CNN- based camera model evaluation pipeline. In Figure 4.2, we assume a similar structure to the
previously presented pipeline in Section 4.2. Our adversarial image generator module takes as input the set of K
patches that have been extracted from the image I that is being analyzed. When presented with new image
patches, our module can work in two different modes.
Figure 4.1: Block diagram of our proposed method.

In the first operation mode, the adversarial image generator module does an untargeted image manipulation, that
is, it does not try to perturb the image patches to produce a specific misclassification class. Instead, we use the
derivative of the loss function of the CNN with respect to the input image patches to add a perturbation to the
images. The derivative is computed using backpropagation with the labels LˆkJ, k ∈ [1, K] that are given by the
CNN detector when it first processes the unmodified image patches. This procedure is known as the fast gradient
sign method (FGSM) [13-14].
In the second operation mode, the adversarial image generator module does a targeted image manipulation. In
this case, we try to perturb the image patches to produce a specific misclassification class LJ, different from the
true real label L that is associated with the analyzed image I and its associated Pk patches. In this mode of
operation, we exploit the 1This is joint work with Dr. Yu Wang, Dr. Luca Bondi, Prof. Paolo Bestagini, and
Prof. Stefano Tubaro forward derivative of a CNN to find an adversarial perturbation that will force the network
to misclassify the image patch into the target class by computing the adversarial saliency map. Starting with an
unmodified image patch, we perturb each feature by a constant offset ϵ. This process is repeated iteratively until
the target misclassification is achieved. This procedure is known as the Jacobian-based saliency map attack
(JSMA) [15].
Its present a detailed overview of both FGSM and JSMA techniques as follows.
4.1. Fast Gradient Sign Method
In, the fast gradient sign method was introduced for generating adversarial examples using the derivative of the
loss function of the CNN with respect to the input feature vector. Given an input feature vector (e.g. an image),
FGSM perturbs each feature in the direction of the gradient by magnitude ϵ, where ϵ is a parameter that
determines the perturbation size. For a network with loss J(Θ, x, y), where Θ represents the CNN predictions
for an input x and y is the correct label of x, the adversarial example is generated as
With small ϵ, it is possible to generate adversarial images that are consistently misclassified by CNNs trained
using the MNIST and CIFAR-10 image classification datasets with a high success rate [16-18].
4.2. Jacobian-Based Saliency Map Attack
In [58], an iterative method for targeted misclassification was proposed. By exploiting the forward derivative
of a CNN, it is possible to find an adversarial perturbation that will force the network to misclassify into a
specific target class. For an input x and a convolutional neural network C, the output for class j is denoted Cj(x).
To achieve an output of target class t, Ct(x) must be increased while the probabilities Cj(x) of all other classes j
/= t decrease, until t = arg maxj Cj(x). This is accomplished by exploiting the adversarial saliency map, which is
defined as
for an input feature i. Because we work with images in this chapter, in our case each input feature i corresponds
to a pixel i in the image input x. Starting with a normal sample x, we locate the pair of pixels {i, j} that maximize
S(x, t)[i] + S(x, t)[j], and perturb each pixel by a constant offset ϵ. This process is repeated iteratively until the
target misclassification is achieved. This method can effectively produce MNIST dataset examples that are
correctly classified by human subjects but misclassified into a specific target class by a CNN with a high
confidence [19-20].
5. RESULT DISCUSSION
The dataset splitting approach suggested in to our issue, we assess our solution in this study. This approach has
been specifically designed for the Dresden Image Dataset, which comprises of 73 devices from 25 distinct
camera model families. Each camera shoots a different quantity of pictures. Every angle offers a different reason
for shooting.
A scene is a grouping of a certain motivation and a specific location. Keeping this criterion in mind, there are 83
total scenes in the dataset. We only take into account camera models with more than one instance since we are
attempting to categorise picture patches at the level of camera model rather than instance level. This results in a
total of 18 camera types (since the only real difference between the Nikon D70 and D70s is their on-device
screen) and around 15,000 photographs.
The dataset, which consists of 18 camera models, is divided into three sets: training DT, validation DV, and
evaluation. DE. DT is once more divided into two equal groups, Dcam T and D ip T. While D ip T is used to
train Mip, Dcam T is used to train the parameters of Mcam.
To prevent overfitting, DV is utilised to determine how many epochs to employ during training. The trained
network is then fairly evaluated using DE on a disjoint collection of pictures.
Avoiding overfitting the data is crucial while training a CNN. Our objective is to identify the camera model from
an image in our dataset, which contains pictures of various scenarios captured by various cameras. It is crucial
that DT and DV differ enough from one another since DV is utilised to prevent overfitting. Additionally, it is
crucial that we test using data that differs from the training set. We take the following actions in order to fulfil
these objectives:
A single instance per camera and a collection of 11 scenarios are used to choose the images for DE.
Images for DT are chosen from 63 distinct situations and extra camera instances.
For DV, images are chosen from the remaining 10 scenes using the same camera instances as for DT.
This separate sets DV and DT in terms of scenes, resulting in strong training.
K = 300 non-overlapping colour patches with a size of 64 by 64 are taken from each image for training,
validation, and testing. There are more than 500 000 patches in the generated dataset Dcam T, which are divided
into 18 classes. According to the Mcam classification findings, D ip T is lowered to 90 000 patches to balance
dependable and unreliable image patches. Finally, there are more than 700 000 and 800 000 patches in DV and
DE, respectively.
5.1. Instructional Techniques
We suggest a two-tiered transfer learning-based strategy called Transfer since the suggested approach relies on a
pretrained network (i.e., Mcam). We also examine the Scratch and Pre-Trained extra tactics for comparison's
purposes. Following, we provide information on each technique in more depth.
Scratch. The most straightforward training method is this one.
The whole two-class architecture M is trained using simply D ip T for training and DV for validation. This might
be regarded as a fundamental training approach. We employ the Adam optimizer with batch size set to 128 and
default parameters as indicated in. Binary-cross entropy loss is the default.
Pre-Trained. This tactic benefits from the potential use of a Mcam that has already been trained. In this instance,
we use the output of SoftMax normalisation to train Mcam for camera model attribution. Validation is done on
DV, while training is done on Dcam T. After Mcam has been trained, we freeze its weights and use Dip T and DV
to train the remaining components of the architecture Mip as a two-class classifier (i.e., reliable vs. non-reliable
patches). The Adam optimizer is used during both training phases to optimise, employing batches of 128 patches
and default values. Our loss function is categorical-cross entropy.
Transfer. This two-tiered training approach aims to fully utilise the suggested architecture's capacity for transfer
learning. The first phase is training Mcam for camera model attribution using Dcam T and DV as datasets, SoftMax
normalisation on its output, and. With default settings, 128 patches per batch, and categorical-cross entropy as
the loss function, Adam is used to optimise this training phase.
The second stage of M training entails freezing all of Mcam's convolutional layers and continuing to train all of
Mcam and Mip's inner product layers using the datasets Dip T and DV. This makes it possible to concurrently learn
the classifier Mip's weights and adapt the feature extraction process in Mcam's last two layers (i.e., ip1 and ip2) to
the classification problem. For this stage, we utilise stochastic gradient descent (SGD) with a fluctuating learning
rate between 5 and 15 10 5 as the optimizer and binary-cross-entropy as the loss metric.
This decision is inspired by exploratory research done in [17] and empirically supported in our study.
Its explain the experimental findings in this section. The first demonstrate how the suggested method can tell
which patches include camera model information and which ones are unsuitable for this job. Then, the
demonstrate how the suggested technique may be used to enhance camera model recognition.
5.2. Patch Reliability
Architecture on CNN. The initial round of tests focused on selecting the network architecture for Mip. To do
this, we used the pre-Trained method to train a variety of architectures over the course of 15 epochs. We chose
all conceivable arrangements of up to six inner product layers (with ReLU activation), each made up of 32, 64,
or 128 neurons, as our layouts. Two neurons are always set as the last layer, followed by softmax. The model
Mip with the best validation accuracy for each tested number of layers was chosen from this experiment, and it
is M2 ip comprised of two inner product layers with 128 and 2 neurons, respectively.
(a) Loss
(b) Accuracy
Figure 5.1: Loss and accuracy curves for the M4 Pre-Trained and Transfer techniques on training
(Dip) and validation (DV) datasets.
composed by three inner product layers with 64, 128 and 2 neurons, respectively.
composed by four inner product layers with 64, 32, 128 and 2 neurons, respectively.
composed by five inner product layers with 64, 32, 64, 128 and 2 neurons, respectively.
composed by six inner product layers with 64, 32, 32, 64, 64 and 2 neurons, respectively.
Loss and accuracy curves on training (Dip T) and validation (DV) datasets using Pre-Trained and Transfer
strategies on .
CNN Architecture
Figure 5.2: The Transfer approach yields the most precise results for any design.
False Positive Rate

Figure 5.3: ROC curves for accurate patch recognition. Solid lines are used to depict the suggested
techniques. Additional baselines are dashed or dotted
Baseline evaluation. We also took into consideration two potential baseline solutions in order to better validate
the suggested strategy.
Table 5.1: Using just the credible patches from the test dataset, one may accurately attribute camera
models
Mip Strategy Patches Accuracy Acc. Delta
Scratch 553 475 0.9009 0.0342
M2 Pre-Trained 618 958 0.9478 0.0811
ip Transfer 637 135 0.9511 0.0844
Scratch 518 228 0.9041 0.0374
M3 Pre-Trained 626 767 0.9520 0.0853
ip Transfer 641 808 0.9554 0.0885
Scratch 562 897 0.8963 0.0296
M4 Pre-Trained 649 515 0.9499 0.0832
ip Transfer 647 998 0.9530 0.0864
Scratch 511 425 0.9045 0.0378
M5 Pre-Trained 648 665 0.9529 0.0862
ip Transfer 651 508 0.9520 0.0853
Scratch 517 386 0.9035 0.0367
M6 Pre-Trained 651 405 0.9501 0.0834
ip Transfer 652 308 0.9541 0.0867
The first involves utilising several supervised classifiers and using the 18-element vector that Mcam returned as
a feature. For this, we trained a gradient boosting classifier (Boosting), a decision tree (Tree), a random forest
(Forest), and a logistic regressor (Logistic). After performing parameter grid-search training on D, we applied z-
score feature normalisation to each approach and chose the model that achieved the best validation accuracy on
DV. On evaluation set DE, accuracy findings for patch reliability were 70.7%, 73.9%, 78.6%, and 81.8%,
respectively. None of them come close to making up the 86% of the suggested answer. The quality-function
described in [18] (Quality- Function) is the second baseline solution that we examined. This function calculates
the suitability of each patch for training Mcam, and it outputs a value between 0 and 1 for each patch. We
decided that a comparison was required for thoroughness even though Quality-Function was not meant to serve
as a test reliability indicator. Figure 3.7 depicts receiver operating characteristic (ROC) curves that were
generated by thresholding our reliability likelihood estimation gk, the soft output of the other classifiers (such as
the logistic regressor, decision tree, etc.), and the quality-function for this purpose.
There are always more chosen patches in E when employing the Transfer approach (bold) than there are in E
overall. Accuracy has risen by more than 8% compared to random patch selection. yielded the value [7]. As
anticipated, using the quality-function described in [7] yields less precise results. In contrast, when trained using
the Transfer technique, the suggested method outperforms every other classifier.
5.3. Attribution of Camera Model
We first examined the suggested method's ability to choose trustworthy patches, and then we looked at how it
affected camera model attribution. For this reason, we present in Table 3.1 the evaluation set findings for the
three training procedures and the five Mip models that were evaluated when a single patch was used to attribute
camera models.
Patches, or the expected total of valid patches.
Accuracy, or the typical result of camera model attribution.
Accuracy Delta, or the improvement in accuracy compared to choosing patches at random rather than
employing patch selection.
These findings show that a greater than 8% improvement in camera model attribution is feasible.
Figure 5.4 displays the confusion matrix findings for assessment data DE randomly picking patches using
Mcam. 87% accuracy on average per patch. Figure 3.9 displays the same outcomes when using M4 to evaluate
just patches that the re deemed credible. The accuracy rises to above 95% in this case. Comparing the two
figures reveals that only the application of trustworthy patches can correct numerous erroneous classifications
outside of the confusion matrix diagonal.
Predicted label
Figure 5.4: Camera model attribution confusion matrix obtained with Mcam on DE without patch
selection
Predicted label
Figure 5.5: Confusion matrix for the camera model's attribution produced using Mcam on DE using
patches chosen.
The suggested technique provides a reliability mask that emphasizes which picture areas are regarded as
dependable in terms of camera attribution in addition to the effect on camera model attribution. This might be
helpful in the future to identify which visual features camera model attribution CNNs value more highly.
Additionally, it may be used in conjunction with localization techniques based on camera model traces to splice
out potentially inaccurate regions from the study.
6. CONCLUSION Proceedings of the IEEE Conference on
The effectiveness of identifying video modifications Computer Vision and Pattern Recognition, pp.
can be significantly increased with appropriate data 770–778, Jun. 2016, Las Vegas, NV. [Online].
and the use of simple machine learning classifiers. Available:
Previous efforts in detecting video manipulation have https://doi.org/10.1109/CVPR.2016.90
primarily focused on examining pixel data to identify [5] Huang, Z. Liu, K. Q. The inberger, and L. van
fabricated content. Our approach involves the der Maaten, “Densely connected convolutional
combination of a random forest and an SVM, trained networks,” Proceedings of the IEEE
on characteristics of Multimedia streams derived from Conference on Computer Vision and Pattern
both genuine and falsified videos. This method has Recognition, Aug. 2016, Honolulu, HI.
proved highly successful in detecting video [Online]. Available:
alterations, even with limited data. Our ongoing https://doi.org/10.1109/CVPR.2017.243
research aims to develop techniques that
automatically sanitize data based on these findings. [6] J. Thies, M. Zollho¨fer, M. Stamminger, C.
Specifically, the plan to remove auxiliary header Theobalt, and M. Nießner, “Face2Face: Real-
information and metadata that could potentially reveal time face capture and reenactment of RGB
sensitive information about the video's source. videos,” Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition,
In future. possible future work could involve pp. 2387–2395, Jun. 2016, Las Vegas, NV.
extending the machine learning approaches used in [Online]. Available:
broadcasting forensics to more complex video https://doi.org/10.1109/CVPR.2016.262
manipulation scenarios. For instance, exploring how
deep learning models such as convolutional neural [7] B. Bayar and M. C. Stamm, “A deep learning
networks (CNNs) can be used to identify more approach to universal image manipulation
sophisticated tampering techniques, such as deepfake detection using a new convolutional layer,”
videos or subtle manipulations that are hard to detect Proceedings of the ACM Workshop on
using traditional approaches. Information Hiding and Multimedia Security,
pp. 5–10, Jun. 2016,
Another area of research could be the development of
real-time video manipulation detection systems that [8] Vigo, Spain. [Online]. Available:
can be used by broadcasters to ensure the integrity of https://doi.org/10.1145/2909827.2930786
live content. This would require exploring how to [9] Xu, H. Z. Wu, and Y. Q. Shi, “Structural design
implement machine learning models with low latency of convolutional neural networks for
and high accuracy on live video streams. steganalysis,” IEEE Signal Processing Letters,
REFERENCES vol. 23, no. 5, pp. 708–712, May 2016.
[1] L. N. Smith, “Cyclical learning rates for [Online]. Available:
training neural networks,” Proceedings of the https://doi.org/10.1109/LSP.2016.2548421
IEEE Winter Conference on Applications of [10] L. Baroffio, L. Bondi, P. Bestagini, and S.
Computer Vision, pp. 464–472, Mar. 2017, Tubaro, “Camera identification with deep
Santa Rosa, CA. [Online]. Available: convolutional networks,” arXiv:1603.01068,
https://doi.org/10.1109/WACV.2017.58 Mar. 2016. [Online]. Available:
[2] K. Liang, J. K. Liu, R. Lu, and D. S. Wong, http://arxiv.org/abs/1603.01068
“Privacy concerns for photo sharing in online [11] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens,
social networks,” IEEE Internet Computing, and Z. Wojna, “Rethinking the inception
vol. 19, no. 2, pp. 58–63, Mar. 2015. [Online]. architecture for computer vision,” Proceedings
Available: of the IEEE Conference on Computer Vision
https://doi.org/10.1109/MIC.2014.107. and Pattern Recognition, pp. 2818–2826, Jun.
[3] L. Bondi, L. Baroffio, D. Gu¨era, P. Bestagini, 2016, Las Vegas, NV. [Online]. Available:
E. J. Delp, and S. Tubaro, “First steps towards https://doi.org/10.1109/CVPR.2016.308
camera model identification with convolutional [12] G. Huang, Z. Liu, K. Q. The inberger, and L.
neural networks,” IEEE Signal Processing van der Maaten, “Densely connected
Letters, vol. 24, no. 3, pp. 259–263, Mar. 2017. convolutional networks,” Proceedings of the
[Online]. IEEE Conference on Computer Vision and
[4] K. He, X. Zhang, S. Ren, and J. Sun, “Deep Pattern Recognition, Aug. 2016, Honolulu, HI.
residual learning for image recognition,”
[Online]. Available: frame rate up-conversion based on residual
https://doi.org/10.1109/CVPR.2017.243 signals,” IEEE Transactions on Circuits and
[13] F. Chollet, “Xception: Deep learning with Systems for Video Technology, vol. 28, no. 7,
depthwise separable convolutions,” pp. 1497– 1512, Jul. 2018. [Online]. Available:
Proceedings of the IEEE Conference on https://doi.org/10.1109/TCSVT.2017.2676162
Computer Vision and Pattern Recognition, pp. [18] R. Raghavendra, K. B. Raja, S. Venkatesh, and
1800–1807, Jul. 2017, Honolulu, HI. [Online]. C. Busch, “Transferable deep-cnn features for
Available: https://doi.org/10. detecting digital and print-scanned morphed
1109/CVPR.2017.195 face images,” Proceedings of the IEEE
[14] G. Wang, J. Cai, and T. Chen, “Recent Conference on Computer Vision and Pattern
advances in convolutional neural networks,” Recognition Workshops, pp. 1822–1830, Jul.
Pattern Recognition, vol. 77, no. C, pp. 354– 2017, Honolulu, HI. [Online]. Available:
377, May 2018. [Online]. Available: https://doi.org/10.1109/CVPRW.2017.228
https://doi.org/10.1016/j.patcog.2017.10.013 [19] P. Zhou, X. Han, V. I. Morariu, and L. S.
[15] S. Suwajanakorn, S. M. Seitz, and I. Davis, “Two-stream neural networks for
Kemelmacher-Shlizerman, “Synthesizing tampered face detection,” Proceedings of the
obama: Learning lip sync from audio,” ACM IEEE Conference on Computer Vision and
Transactions on Graphics, vol. 36, no. 4, pp. Pattern Recognition Workshops, pp. 1831–
95:1–95:13, Jul. 2017. [Online]. Available: 1839, Jul. 2017, Honolulu, HI. [Online].
https: //doi.org/10.1145/3072959.3073640 Available:
https://doi.org/10.1109/CVPRW.2017.229
[16] P. Mullan, D. Cozzolino, L. Verdoliva, and C.
Riess, “Residual-based forensic comparison of [20] L. Bondi, S. Lameri, D. Gu¨era, P. Bestagini, E.
video sequences,” Proceedings of the IEEE J. Delp, and S. Tubaro, “Tampering detection
International Conference on Image Processing, and localization through clustering of camera-
pp. 1507–1511, Sep. 2017, Beijing, China. based cnn features,” Proceedings of the IEEE
[Online]. Available: Conference on Computer Vision and Pattern
https://doi.org/10.1109/ICIP.2017.8296533 Recognition Workshops, pp. 1855–1864, Jul.
2017, Honolulu, HI. [Online]. Available:
[17] X. Ding, G. Yang, R. Li, L. Zhang, Y. Li, and https://doi.org/10.1109/CVPRW.2017.232
X. Sun, “Identification of motion-compensated

Broadcasting Forensics Using Machine Learning Approaches

Uploaded by

Copyright:

Available Formats

Broadcasting Forensics Using Machine Learning Approaches

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Broadcasting Forensics Using Machine Learning Approaches

Uploaded by

Copyright:

Available Formats

International Journal of Trend in Scientific Research and Development (IJTSRD)

Broadcasting Forensics Using Machine Learning Approaches

ABSTRACT How to cite this paper: Amit Kapoor |

Figure 1.2: Deep CNN architecture taxonomy.

Figure 4.1: Block diagram of our proposed method.

False Positive Rate

You might also like