Wasit Journal of Computer and Mathematics Science
Journal Homepage: https://wjcm.uowasit.edu.iq/index.php/WJCM
e-ISSN: 2788-5879 p-ISSN: 2788-5879
Detection of Deep Fake in Face Images Using Deep Learning
Hanady Sabah Abdul kareem1∗ and Mohammed Sahib Mahdi Altaei1
1 Computer Science Department, College of Science, Al-Nahrain University, Baghdad, Iraq
*Corresponding Author: Hanady Sabah Abdul kareem
DOI: https://doi.org/10.31185/wjcm.92
Received: September 2022; Accepted: November 2022; Available online: December 2022
ABSTRACT: Fake images are one of the most widespread phenomena that have a significant influence on our
social life, particularly in the world of politics and celeb. Nowadays, generating fake images has become very easy
due to the powerful yet simple applications in mobile devices that navigate in the social media world and with the
emergence of the Generative Adversarial Network (GAN) that produces images which are indistinguishable to the
human eye. Which makes fake images and fake videos easy to perform, difficult to detect, and fast to spread. As a
result, image processing and artificial intelligence play an important role in solving such issues. Thus, detecting fake
images is a critical problem that must be controlled and to prevent these numerous harmful effects. This research
proposed utilizing the most popular algorithm in deep learning is (Convolution Neural Network) to detect the fake
images.
The first steps includes a preprocessing which start with converting images from RGB to YCbCr color space, after
that entering the Gamma correction. finally extract edge detection by entering the Canny filter on them. After that,
utilizing two different method of detection by applying (Convolution Neural Network with Principal Component
Analysis) and (Convolution Neural Network without Principal Component Analysis) as a classifiers.
The results reveal that the use of CNN with PCA in this research results in acceptable accuracy. In contrast, using
CNN only gave the highest level of accuracy in detecting manipulated images.
Keywords: Deep learning, Deep fake, Generative Adversarial Network, Con- volution Neural Network, Principal
Component Analysis
1. INTRODUCTION
Due to the technological development and the wide spread of programs for deep fakes on digital images with the advent
of Generative Adversarial Network (GAN). Image and video editing are becoming easier and considered a cybercrime due
to the fake information and images can be spread fast and widely in Internet through social media. Thus, deepfake can be
considered as the ability to automatically create, modify, or swap a person’s face in videos and images using algorithms
depending on Deep Learning technology, that is one of the phenomena that is expanding quickly. It is feasible to produce
top-notch results by developing new multimedia materials that are difficult for the human eye to distinguish between real
and fake. This term "Deepfake" refers to all multimedia items that have been synthetically modified or produced using
generative machine learning models [1]. DeepFake is concerning because it combines a high level of authenticity, quick
evolution and pervasiveness [2]. On Reddit, DeepFake has been widely utilized as of November 2017. due to the United
States widespread manufacturing of pornographic videos that has gained a solid online reputation and attracted interest of
the people from all walks of life. In January DeepFake was officially used in an application in 2018.As a result DeepFake
proliferation was accelerated. The object of face swapping has also grown in popularity from celebrities and politicians
to students, friends and coworkers As a result, people’s panic were naturally prompted. The main IT companies have also
https://wjcm.uowasit.edu.iq/index.php/wjcm
Hanady Sabah Abdul kareem , Wasit Journal of Computer and Mathematics Science, Vol. 1 No. 4 (2022) p. 60-71
begun to collaborate action with the academic community to avoid additional detrimental impact on the fierce discussion
about whether and how to control Deepfake technology [3].
With the continuous enhancement of the computing power of computer, the ongoing reduction of hardware price
and the high integration of deep learning tools like tensorflow [4] and keras [5], technicians with specific professional
backgrounds can produce high-quality faked images and videos consistent with the real distribution of data through the
use of convolution automatic encoder [6] and Generative adversarial networks (GAN) [7]. Deepfake image as display in
Figure 1
FIGURE 1. Several illustrations of DeepFake: (a) Obama, made by Buzzfeed in cooperation with Monkeypaw Studios; (b)
Mark Zuckerberg, made by artists Bill Posters and Daniel Howe in association with advertising company canny ; (c) Matteo
Renzi„ made by “Striscia la Notizia [8].
2. RELATED WORK
1. Y. Wang et al in 2021 presented two algorithms for detection of fakeface images. The first approach is the Local
Binary Pattern (LBP)-Net using global texture features used to detect fakefaces. The second method ensemble model
constructed from five models including LBP-Net, Gram-Net, ResNet and two models utilizing InceptionResnetV1
pre-trained on Casia-Webfaceare and vggface2. Results of detecting fakeface images by several image augmentation
such as downsample (66.32%), brightness (81.09% ), Solarize ( 75.04%) ,Contrast ( 85.42%) and color (91.06%)
when using “140K Real and Fake Faces” [9].
2. M. Taeb et al in 2022 compared the most popular state-of-the-art face-detection classifiers such as CNN, VGG19,
and DenseNet-121 using an enhanced actual and fakeface dataset. Data augmentation is a technique for improving
performance with conserving computing resources. When compared to other studied models, early findings show
that VGG19 has the best performance and accuracy of 95%. When using “140K Real and Fake Faces” [10].
3. S. Tariq and et al. in 2018 proposed an ensemble of three different convolution neural network with different layers
as a classifier after performing pre-processing represented by face cropping and noise filtering methods. Finally a
fully automated end-to-end fakeface detection pipeline developed to be focused on image content with only RGB
channel information in order to recognize GAN-created face images with 94% AUROC score (Area Under the
Receiver Operating Characteristic curve) and recognize human-created fake face images with 74.9 % AUROC
score. Where, applied on CelebA and PGGAN dataset images. [11].
4. X. Chang et al in 2020 improved VGG network termed NA-VGG, it is used to detect deepfake face images, which
was based on image noise and augmentation of image. The SRM filter layer is utilized to highlight the noise
features of the image and after that the image noise is acquired as the network’s input. Second, the image noise
map is augmented to make the face features appear weaker. Finally, the augmented noise of images are fed into the
network, which is trained and used to determine whether or not the image is fake. The obtained accuracy is 85.7 %
when using the Celeb-DF dataset [12].
61
Hanady Sabah Abdul kareem , Wasit Journal of Computer and Mathematics Science, Vol. 1 No. 4 (2022) p. 60-71
3. THE WORKFLOW OF THE MODEL
This project has been performed in five steps. The broad discussion of these stages is described here. The first stage is
choosing the suitable real and fake images dataset from kaggel.com and preprocessing the dataset. following that, PCA
applying for selecting images features after splitting the dataset by using cross-validation (hold-out) (80:20). The next
stage is using (CNN) classifiers to classify the dataset. Finally evaluate performance of model using various metrics such
as (accuracy, precision, recall, and F1-score). as showed in Figure 2.
FIGURE 2. Work design step of fake images detection.
The Proposed Deep Fake Detection
3.1 DATASET
Deepfake detection systems often use binary classifiers to group information into fake and real classes. This strategy
needs a great quantity of high-quality authentic and manipulated data to train the models of classification. The dataset was
taken from kaggle website. This dataset contains of all 70k real faces from of the Flickr dataset gathered by Nvidia, in
addition 70k fake faces picked from the Bojan’s 1 Million fake faces (produced using StyleGAN).
In this dataset, combined both dataset, scaled all of the images to 256px, and divided the data into three sets: train,
validation and test set. also some CSV files available for convenience [13]. In this study, just two features (images and
labels) are utilized to detect fake image classifiers. Label one represents fake images, whereas label zero represents true
images.
FIGURE 3. Samples of the dataset’s real images
62
Hanady Sabah Abdul kareem , Wasit Journal of Computer and Mathematics Science, Vol. 1 No. 4 (2022) p. 60-71
FIGURE 4. Samples of the dataset’s fake images
3.2 PRE-PROCESSING STEPS
In this section the results of the pre-processing of the images were reviewed, where six images were taken randomly
from the dataset that was used in the proposed system.The first three represent the real images and the second three
represent the fake ones, where the figure shows the results before and after the pre-processing. The first column represents
the original images with RGB color space, while the second column represents the images conversion from RGB into
YCbCr color space, the third column represents the images after entering the Gamma correction and finally the fourth
column represents the images after entering the Canny filter on them. As displayed in Figure 5.
63
Hanady Sabah Abdul kareem , Wasit Journal of Computer and Mathematics Science, Vol. 1 No. 4 (2022) p. 60-71
FIGURE 5. Resulted of a preprocessing sample image from 140k Real and Fake Faces .
3.3 PRINCIPAL COMPONENT ANALYSIS (PCA)
Principal Component Analysis (PCA) is one of the most widely statistical approaches used for feature selection. It has
several uses in picture compression, text classification, and face recognition [14]. This is a frequently used method for
reducing the dimensionality of a feature collection via a linear transformation. The main goal of PCA is to reduce the
original variables to subset of variables by calculating the highest relationship of the original variables [15]. Although the
resultant dataset is reduced, but the original data set’s features are still retained and removed the redundancy of informa-
tion [16] and [17]. The number of features in the new dataset may be equal to or lower than that in the original dataset.
The principal components are computed by using the covariance matrix. The ability discriminative of the classifiers can
be improved by using PCA .
The following steps summarize the process [18].
1. Let the training set of images {X1, X2, X3. . . XN}. Calculation of mean value of an image with equation:
n
− 1X
x= xi (1)
n i=1
Information:
xi = data variable
n = numbers of data
Calculation of the mean value to reduce the dimension to be calculated in the next process.
2. Compute the covariance matrix to depict the scatter degree of all feature vectors connected to the average vector.
The covariance matrix C is defined as follows:
n
1X − −
c= (xi − x)(xi − x)T (2)
n i=1
3. Calculate the Eigenvectors and eigenvalues from covariance matrix
CV = lV (3)
Where V is the set of eigenvectors correlated with its eigenvalue .
4. Sorts the eigenvector and eigenvalues from high to low based on the order of eigenvalues.
The main component k is the eigenvector corresponding to the largest eigenvalue.
The main component k of the vector x is observed using the equation:
−
wi = viT (xi − x) (4)
64
Hanady Sabah Abdul kareem , Wasit Journal of Computer and Mathematics Science, Vol. 1 No. 4 (2022) p. 60-71
3.4 CLASSIFIER
Convolutional Neural Networks is a type of Artificial Neural Net-work (ANN), also known as Conv-Nets or CNN,
That is one of the most effective deep designs for classifying image data and many other applications, including audio
recognition, object detection, medical image analysis and natural language processing [19]. The deep architecture of the
network produces hierarchical feature extraction, wherein the trained filters of the first layer can be seen as a set of color
dots or edges, of the second layer as some forms, the filter of the next layer may learn part of objects and the final layers
filters may be able to recognize the objects. [20].
In the proposed method the design is carried out in two stages. Which starts first by training the model on the dataset
consisting of real images and fake images. As for the second stage which is the testing stage, it is easy to distinguish the
test image whether it is fake or real. One of the most common algorithms used in classification is a convolutional neural
network (CNN) architecture by created a model from scratch with using six blocks in the CNN architecture. In each block
utilize Conv2D with kernel size=5, Max Pooling, batch normalization=64, model = Sequential, dropout, activation =
Relu, padding = same, epsilon=0.001, epochs = 100 One iteration on each training data set is represented as an epoch,
verbose = 1, shuffle=True and to optimize the network using Adam optimizer = 0.001and a learning rate 0.0001. As
discussed in Table 1.
Once obtaining the features (components) the features are fed to Convolution Neural Network (CNN) layer of the deep
65
Hanady Sabah Abdul kareem , Wasit Journal of Computer and Mathematics Science, Vol. 1 No. 4 (2022) p. 60-71
Table 1. CNN architectures for deep fake detection of face image. Each row represents a block in the architecture. (Note:
Conv=Conv2D, BN=BatchNormalization, DO=Dropout, MP=MaxPooling, F=Flatten, GAP=GlobalAveragePooling,
D=Dense & S=Sigmoid)
Convolution Neural Network
Conv-ReLU-Padding-MP-BN
Conv-ReLU-Padding-MP-BN-DO
Conv-ReLU-Padding-MP-BN-DO
Conv-ReLU-Padding-MP-BN-DO
Conv-ReLU-Padding-MP-BN-DO
Conv-ReLU-Padding-MP-BN-DO
GAP-F-D-ReLU-DO
D- ReLU-DO-D-S
learning model, which further selects the useful features by using their filters. The selected features are supplied to the
max-pooling layer to choose the features that have the highest importance value throughout the computation. Gradients
will be calculated, and the network’s weights will automatically adjust.. We report our detection accuracy in result section.
3.5 EVALUATION METRICS
In order to evaluate the accuracy of algorithm’s classification in detecting fake images, a number of evaluation metrics
were used. The most commonly used measuring metric (Confusion Matrix) for detecting fake images has been applied in
this part. It is possible to specify the metrics that the confusion matrix will use by defining this as a classification task. has
as below [21].
Accuracy = T P+TTN+FP+FN
P+FP
[22]
Precision = T P+FP [23]
TP
Recall = T P+FNTP
[24]
F1 Score = Precision+Recal [25]
2∗ Precision∗Recall
where TP represents (True Positive) and TN represent (True Negative). Moreover, FP is False positive, and FN is
False-negative. As discussed in Table 2.
Table 2. Parameters of evaluation metrics
Parameters Description
TP True Positive The number of correctly classified records
TN True Negative The number of the correctly rejected of records that have been classified
FP False Positive The number of incorrectly classified records
FN False Negative The number of the incorrectly rejected of records that have been classified
These measures are often used in a sequence of machine learning algorithms to evaluate the performance of a classifier
from several estimations. Specifically, the accuracy measure, which quantifies the similarity between predicted and actual
fake images. Precision refers to the measuring of the fraction of the found fake images, that has been classified as fake,
addressing the critical issue of the fake images classification. However, because to the dataset of fake images is typically
skewed, good precision may be achieved by making a lower number of optimistic predictions, hence recall is used to
assess sensitivity, or the fraction of annotated fake articles predicted as fake. It should be highlighted that higher values
refers to better Recall, Precision, and Accuracy [26].Finally F1 score": Another term for it is the F Score or F Measure.
The F1 score is a great balance of P and R [25].
4. RESULTS AND DISCUSSION
According on the classification results showed that the accuracy of the model with pre-processing is CNN only without
PCA and CNN with PCA classifier is 63.86% and 74.26%, respectively. Table 3, Table 4, Table 5, Table 6, Table 7 and
Table 8 displays the resulted of confusion matrix with TP, FP, TN and FN values. Our experimental findings without
preprocessing stages show that CNN only without PCA achieves 93.16% and CNN with PCA achieves 90.76%. Finally,
an additional experiment was conducted by increasing the number of samples entering the CNN network ( Training
and Testing) which gave the highest accuracy results in image classification, so that the classification accuracy reached
98.04%.
Accuracy =63.86
66
Hanady Sabah Abdul kareem , Wasit Journal of Computer and Mathematics Science, Vol. 1 No. 4 (2022) p. 60-71
(a) CNN without PCA with preprocessing
FIGURE 6. The curves of the validation accuracy and loss during the training stage for the proposed CNN without PCA with
preprocessing.
Table 3. Confusion matrix of the detection of deep fake on face image using CNN only without PCA with preprocessing
Positive Negative
Positive 1064 436
Negative 648 852
Precision=70.93
Recall=62.14
F1 Score=66.24
(b) CNN with PCA with preprocessing
FIGURE 7. The curves of the validation accuracy and loss during the training stage for the proposed CNN with PCA with
preprocessing.
Table 4. Confusion matrix of the detection of deep fake on face image using CNN with PCA with preprocessing
Positive Negative
Positive 1144 356
Negative 416 1084
Accuracy =74.26
Precision=76.26
Recall=73.33
F1 Score=74.76
Accuracy =93.16
Precision=94.26
67
Hanady Sabah Abdul kareem , Wasit Journal of Computer and Mathematics Science, Vol. 1 No. 4 (2022) p. 60-71
(c) CNN only without PCA on colored image
FIGURE 8. The curves of the validation accuracy and loss during the training stage for the proposed CNN only.
Table 5. Confusion matrix of the detection of deep fake on face image using CNN only without PCA on colored image
Positive Negative
Positive 1414 86
Negative 119 1381
Recall=92.23
F1 Score=93.23
(d) CNN with PCA on colored image
FIGURE 9. The curves of the validation accuracy and loss during the training stage for the proposed CNN with PCA.
Table 6. Confusion matrix of the detection of deep fake on face image using CNN with PCA on colored image
Positive Negative
Positive 1373 127
Negative 150 1350
Accuracy =90.76
Precision=91.53
Recall=90.15
F1 Score=90.83
Table 7. Confusion matrix of the detection of deep fake on face image using CNN only without PCA on colored image
Positive Negative
Positive 7833 167
Negative 146 7854
68
Hanady Sabah Abdul kareem , Wasit Journal of Computer and Mathematics Science, Vol. 1 No. 4 (2022) p. 60-71
(e) CNN only without PCA on colored image
FIGURE 10. The curves of the validation accuracy and loss during the training stage for the proposed CNN.
Accuracy =98.04
Precision=97.91
Recall=98.17
F1 Score=98.03
(f) CNN with PCA on gray image
FIGURE 11. proposed CNN with PCA on gray image.
Table 8. Confusion matrix of the detection of deep fake on face image using CNN with PCA on gray image
Positive Negative
Positive 1304 196
Negative 209 1291
Accuracy =86.5
Precision=86.93
Recall=86.18
F1 Score=86.55
In this section, the results were reviewed for implementing the proposed system and the results were as shown in the
explanation below.
The research aims to reveal the manipulation that may be present in the images of faces, and therefore it was necessary
for us to do a test stating that preserving the image entered into the system for the purpose of testing it and detecting the
presence of manipulation in it first. As a result of the initial processing of the image, it may lead to the burial of some
traces of manipulation in the image.
So we tested the image by using the technique involved the use of deep learning technology. In this case, we used
successive steps for a simple preprocessing (determining the size of the image).
We clarified the image and reveal its edges before entering the tamper detection system.
69
Hanady Sabah Abdul kareem , Wasit Journal of Computer and Mathematics Science, Vol. 1 No. 4 (2022) p. 60-71
We found that any preprocessing of the image, led to undesirable results compared to the results obtained without
preprocessing.
Where the network of detection of manipulation in faces using the CNN classifier with the preprocessing and the use
of canny Detection to detect edges was about 63.86
An additional experiment was used by adding the PCA method after the stage of preprocessing, which produced
detection rates of 74.26% for the CNN classifier.
This indicates that the use of the PCA as a feature selector has improved the results of detection by a small percentage,
but it is not high, because the work of the PCA depends on converting the data of the raw image entered into the
detection system into components containing more information that spreads downward from the first component to the
last component. The use of a limited set of component. Since the PCA was used in the form of a feature selector, it relied
on the first components in the classification process by CNN, leaving the last components that contain information but are
very weak and represent a burden on the classification process.
High classification rates and proof that any pre-processing on the image leads to changing or erasing the traces of
manipulation encouraged the use of classification techniques approved by CNN, which is CNN with the PCA and without
pre-processing, the detection rate of manipulation reached 90.76%., while detection rates using CNN only reached to
93.43. A new attempt was also used by inserting the CNN classifier with PCA on a gray scale image and the results were
86.5%.
Finally, CNN classifiers were used to detect forgery and applied to the images directly without using PCA and without
any pre-processing, but with more data to see if the network is affected by the number of data that is trained on, and the
classification results were 98.04% for the CNN classifier, which is the highest accuracy result in image classification It
reached by the proposed system
This indicates that the use of PCA in the proposed system affected the accuracy of the results, and that the process of
inserting direct images into the CNN classifier has greatly benefited in the process of detecting fraud, as CNN needs a lot
of information for the purpose of training and learning from it, regardless of the percentage of information contained in it.
Pictures. CNN also made use of the images as more raw data. From the case that the PCA application has, that is, it made
more use of the images than the components
5. CONCLUSION
The remarkable development of artificial intelligence and with the efficiency of the GAN in generating fake images
that are closer to reality, it was necessary to find an efficient way to detect fake images. the final result of our research
was to confront the phenomenon of Deepfake offered a detection model for the fake images by utilizing the most popular
algorithm in deep learning is (Convolution Neural Network) to detect the fake images. The preprocessing steps start
with converting images from RGB to YCbCr color space, after that entering the Gamma correction. finally extract edge
detection by entering the Canny filter on them. After that, utilizing two different method of detection by applying (CNN
with PCA) and (CNN without PCA) as a classifiers.
The achieved result is better than the listed related work, so utilizing this method enhances the accuracy of classification
. From the above results, we conclude the following:
1. The CNN only is better than a CNN with PCA in the classification accuracy of fake images dataset.
2. CNN is more suitable for large datasets because the network efficiency increases with more data, that is, it gives
better results as the training data increases
3. Preprocessing steps when using our dataset give worse results. These steps had a huge impact on decreasing
classification accuracy.
4. The type of data used has a significant impact on the categorization accuracy of this work.
FUNDING
None
ACKNOWLEDGEMENT
None
70
Hanady Sabah Abdul kareem , Wasit Journal of Computer and Mathematics Science, Vol. 1 No. 4 (2022) p. 60-71
CONFLICTS OF INTEREST
The author declares no conflict of interest.
REFERENCES
[1] L. Guarnera, O. Giudice, and S. Battiato, “Deepfake detection by analyzing convolutional traces,” Proceedings of the IEEE/CVF conference on
computer vision and pattern recognition workshops, pp. 666–667, 2020.
[2] L. S. Wang, 2019.
[3] X. Chang, J. Wu, T. Yang, and G. Feng, 2020.
[4] M. Abadi, P. Barham, J. Chen, Z. Chen, A. Davis, J. Dean, and M. Devin, “{TensorFlow}: a system for {Large-Scale} machine learning,” 12th
USENIX symposium on operating systems design and implementation (OSDI 16), pp. 265–283, 2016.
[5] F. Chollet, 2015.
[6] A. Tewari, M. Zollhofer, H. Kim, P. Garrido, F. Bernard, P. Perez, and C. Theobalt, “Mofa: Model-based deep convolutional face autoencoder
for unsupervised monocular reconstruction,” Proceedings of the IEEE International Conference on Computer Vision Workshops, pp. 1274–1283,
2017.
[7] G. Antipov, M. Baccouche, and J.-L. Dugelay, “Face aging with conditional generative adversarial networks,” 2017 IEEE international conference
on image processing (ICIP), pp. 2089–2093, 2017.
[8] L. Guarnera, O. Giudice, C. Nastasi, and S. Battiato, “Preliminary forensics analysis of deepfake images,” 2020 AEIT international annual
conference (AEIT), pp. 1–6, 2020.
[9] Y. Wang, V. Zarghami, and S. Cui, “Fake Face Detection using Local Binary Pattern and Ensemble Modeling,” 2021 IEEE International Conference
on Image Processing (ICIP), pp. 3917–3921, 2021.
[10] M. Taeb and H. Chi, “Comparison of Deepfake Detection Techniques through Deep Learning,” Journal of Cybersecurity and Privacy, vol. 2, no. 1,
pp. 2022–2022.
[11] S. Tariq, S. Lee, H. Kim, Y. Shin, and S. S. Woo, “Detecting both machine and human created fake face images in the wild,” Proceedings of the
2nd international workshop on multimedia privacy and security, pp. 81–87, 2018.
[12] X. Chang, J. Wu, T. Yang, and G. Feng, 2020.
[13]
[14] A. I. Taloba, A. Dalia, Eisa, S. I. Safaa, and Ismail, 2018.
[15] Karamizadeh, S. M. Sasan, A. A. Abdullah, M. Manaf, A. Zamani, and Hooman, “An overview of principal component analysis,” Journal of
Signal and Information Processing, vol. 4, 2020.
[16] M. Ahmad, A. M. Khan, J. A. Brown, S. Protasov, and A. M. Khattak, “Gait fingerprinting-based user identification on smartphones,” 2016
International Joint Conference on Neural Networks (IJCNN), pp. 3060–3067, 2016.
[17] M. Ahmad, D. I. Ul, Q. Haq, M. Mushtaq, and Sohaib, “A new statistical approach for band clustering and band selection using K-means
clustering,” Int. J. Eng. Technol, vol. 3, no. 6, 2011.
[18] V. Arjun, R. R. Mane, V. Manza, and Karbhari, “Human face recognition using superior principal component analysis (SPCA),” International
Journal of Computer Theory and Engineering, vol. 2, no. 5, 2010.
[19] B. Traore, B. Boubacar, F. Kamsu-Foguem, and Tangara, “Deep convolution neural network for image recognition,” Ecological Informatics,
vol. 48, 2018.
[20] N. Aloysius and M. Geetha, “A review on deep convolutional neural networks,” 2017 international conference on communication and signal
processing (ICCSP), pp. 588–0592, 2017.
[21] J. Ali, R. Khan, N. Ahmad, and I. Maqsood, “Random forests and decision trees,” International Journal of Computer Science Issues (IJCSI),
vol. 9, no. 5, 2012.
[22] P. Shahane and D. Gore, “Detection of Fake Profiles on Twitter using Random Forest & Deep Convolutional Neural Network,” International
Journal of Management, 2019.
[23] H. Allcott and M. Gentzkow, “Social media and fake news in the 2016 election,” Journal of economic perspectives, vol. 31, no. 2, 2017.
[24] S. Gilda, “Notice of violation of IEEE publication principles: Evaluating machine learning algorithms for fake news detection,” 2017 IEEE 15th
student conference on research and development (SCOReD), pp. 110–115, 2017.
[25] J. Han, M. Kamber, and J. Pei, “Data mining concepts and techniques third edition,” The Morgan Kaufmann Series in Data Management Systems,
vol. 5, pp. 83–124, 2011.
[26] K. Shu, A. Sliva, S. Wang, J. Tang, and H. Liu, “Fake news detection on social media: A data mining perspective,” ACM SIGKDD explorations
newsletter, vol. 19, no. 1, 2017.
71
