Fake News Classification using Transfer

ϮϬϮϯ/ŶƚĞƌŶĂƚŝŽŶĂůŽŶĨĞƌĞŶĐĞŽŶƌƚŝĨŝĐŝĂů/ŶƚĞůůŝŐĞŶĐĞĂŶĚ<ŶŽǁůĞĚŐĞŝƐĐŽǀĞƌLJŝŶŽŶĐƵƌƌĞŶƚŶŐŝŶĞĞƌŝŶŐ;/KE&ͿͮϵϳϵͲϴͲϯϱϬϯͲϯϰϯϲͲϰͬϮϯͬΨϯϭ͘ϬϬΞϮϬϮϯ/ͮK/͗ϭϬ͘ϭϭϬϵͬ/KE&ϱϳϭϮϵ͘ϮϬϮϯ͘ϭϬϬϴϯϲϳϴ

Learning
Muthu Lakshmi.V1, K. Vijayakumar2, Suthanthira Devi.P1, Rajin Gangadharan2 and Mr. D. Suresh3
1
Department of Information Technology, St. Joseph’s College of Engineering, OMR, Chennai.
2
Professor, Department of Computer Science and Engineering St. Joseph’s Institute of Technology, OMR, Chennai.

E-mail : [email protected] [email protected] suthanthiradevip@ stjosephs.ac.in

[email protected] [email protected]

Abstract-The rising complexity of information individuals and entire civilizations. People may be
communication technology has greatly affected misled and adopt a false perspective as a result of
communication through conventional broadcast media fake news.People's perceptions of real facts can be
over the past decade. Smartphone applications are affected by counterfeit reports. The widespread
increasingly emasculating the new socio-economic
dissemination of false information has the potential
broadcasting environment. The trend is the same in
the workplace, at home and in recreation. Social
to undermine the validity of the entire reporting
networking has stolen the game and is increasingly environment.
shifting to another age, the era of "digital
relationships," in which conventional interpersonal Therefore, it's important to recognise fake
social interactions are replaced by mobile devices and information on social networking sites. The
social networks. The consequences of such false deliberate printing of false stories to misinform
information promoted by miscreants and apologists clients makes it challenging to identify news that is
for social media are far-reaching because it has factual. The investigation of additional information
resulted in scandals in households, communities, derived from various viewpoints is typical and
partnerships, organizations, and culture as a whole.
necessary for the creation of an effective and
The purpose of this paper is to lead to the eradication
of counterfeit media by the use of technology. In this
practical false report detection system.
article, we proposed and built a model that
incorporates neural networks to identify and eradicate Due to the rising trend of online misinformation,
false phrases posted to social media networks and web which is primarily found in medium outlets such
forums. Also, we compared our work Elmo VNetwith social networking sites, supply, report blogs, and
current state-of the-art models. The experimental online correspondents, the detection of bogus reports
results demonstrated that the proposed Elmo has recently drawn a growing interest from the
VNetmodel have better accuracy rate than the existing general public as well as researchers.
models.
Some approaches provide developers with a great
Index Terms: VNet, Social media, Social networks,
Convolutional Neural networks, LSTM, Recurrent deal of assurance when building systems that can
Neural Network, Elmo Embedding.; unintentionally detect fake reports. However, since
mock-up is required for the summary of studies, it
I.INTRODUCTION can be difficult to identify fraudulent data. A
comparison between true and false information is
Fake news has become one of the greatest issues in done in order to classify a report as false.
our era. It has distorted both online and offline
debate, and one can even go as far as to suggest that, Additionally, the task of comparing projected news
to date, fake news poses a clear and present crisis to with the real piece is a difficult one due to its bias
the community. "False reports" refers to unfounded and estimation. For the classification of the attitude
rumours or misinformation, including propaganda, in the middle of the editorial corpse and caption,
that is spread through traditional media like numerous experiments are conducted for the
publication and television as well as through establishment of a baseline with certain classic
unconventional media like social networking machine learning approaches, and then comparisons
sites.False news have the potential to harm both are made between these findings to the typical deep

1
Authorized licensed use limited to: Malardalen University. Downloaded on April 14,2024 at 22:02:00 UTC from IEEE Xplore. Restrictions apply.
networks. may be created fast and easily to spread false
information. The extraction of various characteristics
II.RELATED WORK from fake news information, such as content based
on expertise and style, is a key component of
Social media is a global forum not only for sharing content-focused approaches to news. User social
opinions but also facts, gossip and fake news. commitments will be used as supplemental
Because tweets are not strictly verified, people tend information in social context-based techniques,
to post fake news to their own advantage. To which are intended to help identify bogus news. [8].
identify fake news content the traditional methods Aggarwal et al., together with URL and WHOIS
that existed are all related to text embedding they are apps, developed an automated real-time phishing
bag of words, inverse document frequency (TF IDF), detector on Twitter using random forest
transform frequency, word2vec, Glove etc. The classification on certain Twitter fields and obtained
authorswere presented with a dataset of some rows an outstanding 92.5 per cent accuracy. One of the
and columns with unique material for any study, major contributions to their work was the Chrome
which can be broken into any ratio for training and Extension for real-time phishing detection that they
testing purposes for classification. The partitioned created and made available on Twitter. [12]. Cook et
training set consisting of data and sentences is stored al. involved in investigating the online bogus
in a separate vocabulary for false news political personas in Australia. He derived a nine-
identification, from which the sentences are way test including: entropy test, spam and miscreant
converged as tokens and placed in index vocabulary. test, account properties, account creation date (if
For a certainpredetermined value, the data points are close election date), pre-election inactivity, Amid
fixed. During classification methodology the user alignments, mass retweets on specific policy days
searches for a data point, if it is found in the and times, and analysis of discrimination (combining
vocabulary it is reconstructed as a vector. This entropy, spam and account, inactivity, alignment and
methodology of finding the data points in mass retweets)[13]. Allcott et al. wrote that the
vocabulary and converting it into vector format is dissemination of falsified news was wherein Donald
called embedding neural network. The drawbacks Trump would be elected as the president of the
will arise among the existing approaches in the United States during the election in the year 2016. A
Recurrent Neural Network if there are broad data-gathering survey was conducted. The collected
sentences that have to be fed for classification into data showed that if there were not so many false
the forward and backward pass. Fette et al. used news about him, Trump would not have won. The
machine learning to identify an email as phishing or websites that distribute fake news appear to be short-
not using features such as URL era, number of dots lived and the key reason for creating fake news was
in URL and HTML email content thus achieving a to raise money from advertisements when news goes
high 99.5 percent accuracy [2]. Bajaj et al. used viral and for other political reasons. Mostly linear
Two-Layer Feed forward Neural Network, Recurrent regression and prediction were used [14].
Neural Network, Convolution Neural Network, Benevenuto et al. have classified real YouTube users
Long-Short Term Memories, and Gated Recurrent as spammers, promoters, and legitimate users. To
Units and found that the GRUs RNN architecture detect promoters and spammers, techniques such as
outperformed one with LSTM cells to detect fake supervised machine learning algorithms were used;
news. The feed forward network performed well as a greater accuracy was achieved for detecting
classifier, but the convolution network did not promoters; algorithms were less efficient for
perform well [5]. Shao et al. explored the detecting spammers. [16]. The aim of Pisarevskaya
correlation between the timing of the dissemination et al. was to reveal false and truthful news using
of false information and the use of a public twitter markers of different linguistic levels. They used
dataset. The dataset was developed by compiling POS tags, word length, sentimental terms, and
tweets from false news and fact-checking websites punctuation at the lexical level Ghosh et al.
with URLs. This study confirms that the most described social agriculture on Facebook, and also
popular way to share platform news stories that limit suggested a technique for combating farming
user expression is through URLs (such as the 140 linkages [17]. Chen et al. said there is a blurring of
character limit imposed by Twitter) [10]. According the line between traditional media advertising and
to Shu et al., fraudulent profiles, like Facebook bots, content created by users. By formulating it as a
query, the rumour is portrayed in the headline as fact

2
Authorized licensed use limited to: Malardalen University. Downloaded on April 14,2024 at 22:02:00 UTC from IEEE Xplore. Restrictions apply.
and the advertised material is presented as native content will be given as another input at the
content. [20]. same time and in the output layer the fake/real
probability will be given.
III. PROPOSED WORK iii) The Siamese net that finds the actual relation
between the two inputs with respect to the
In the proposed work, analysis of different model on output value.
fake news classification is carried out, there are iv) In the Siamese net, we used three state of the art
Siamese CNN, Siamese RNN, Siamese LSTM, layers
Elmo VNet (proposed model). 1. 1D CNN
2. RNN
3. LSTM

B.1D CNN Siamese Network

In Siamese Neural network we use 1D CNN layer
instead of dense layer is called as 1D CNN Siamese
Network. This network has the following
components as shown in Fig 2:
1. Headline
Fig. 1 Siamese Neural Network 2. Actual Content of the news

Siamese Neural Networks

The Siamese neural network is an artificial neural
network that collaborates on two different input
vectors while measuring comparable output vectors
using the same weights. Figure 2. 1D CNN Siamese Network
One of the output vectors is always pre-calculated, Convolutional Layer
thereby creating a baseline that matches the other Convolution is a function that expresses how the
output vector with it. A Siamese network may look shape of one is changed by the other by deriving it
like as stated in Fig 1. from two supplied functions using integration.To
arrive at a better understanding, we could take a look
Siamese Network at the equation of convolution:
Siamese networks' main premise is that they can
acquire useful information descriptors that can be ∞
ሺ݂ ‫݃ כ‬ሻሺ‫ݐ‬ሻ=‫׬‬െ∞ ݂ሺܶሻ݃ሺ‫ ݐ‬െ ܶሻ݀‫ݎ‬ (1)
applied to further contrast the inputs of the
corresponding sub networks. As a result, in this
This architecture is based on a more intuitive
scenario, sub-networks are often created by entirely
approach to the idea of convolution neural networks,
linked layers, CNN-subnetworked picture data, or
not a quantitative or strictly scientific one.
even sequential data like words or time signals. Any
type of input is possible, including numerical
data.Siamese Neural networks in Fake news 1D CNN layer
The convolution layer is the fundamental unit of a
classification:
Convolutional Neural Network. The convolution
layer is simply a kernel function whose parameters
i) The art of filtering falsified news and factual
are trainable using some optimization strategy.
information is a very critical issue for the public
These kernels can also be extended to accommodate
today. The fake news classification is the task of
multiple channels of information making it possible
identifying the fake/real relationship between the
for CNNs to extend the number of channels as we go
News headline and the actual whole context of
deeper into the network. The output of a convolution
the news.
unit/kernel is a two-dimensional matrix. To create
ii) In the Siamese neural network, the news
multichannel intermediate layers, multiple kernels
headline is given as one input and the actual
can be used on the same input the outputs of each

3
Authorized licensed use limited to: Malardalen University. Downloaded on April 14,2024 at 22:02:00 UTC from IEEE Xplore. Restrictions apply.
kernel can be concatenated to obtain a multi-channel formula for the softmax function is:
tensor. ݂ ݁‫݆ ݖ‬ (2)
݆ ሺ‫ݖ‬ሻൌσ ‫ݖ‬
݇݁ ݇
Max pooling layer:
To reduce the complexity of the performance and A cross-entropy loss of 0 for the model, means that
avoid over fitting of the data, a pooling layer is also the model is perfect. The cross-entropy loss
used after a CNN layer. penalizes higher variances between ground truth and
predicted value in a disproportionate manner in
ReLU Layer relative comparison to lower variances between
ReLU (Rectified Linear Unit) is a non-linear ground truth and predicted value. This allows the
activation function that has a linear constant slope model to easily converge.
for any positive input, but has a zero slope for every ‫ܮ‬ ݁ ݂‫݅ݕ‬
ሺ͵ሻ
݅ൌെ Ž‘‰ ቆ ቇ
negative input. Basically, if the input is a vector, the σ݆ ݁ ݂‫݅ݕ‬

ReLU function removes any negative value in that ‫ܪ‬ሺ‫݌‬ǡ ‫ݍ‬ሻ ൌ  െ ෍ ‫݌‬ሺ‫ݔ‬ሻ Ž‘‰ ‫ݍ‬ሺ‫ݔ‬ሻሺͶሻ
vector and replaces them with a 0, and doesn’t
‫ݔ‬
modifies the positive values in the vector. This
function helps to approximate a decision function C. Siamese RNN for fake news classification
without affecting the overall properties of the output In Siamese Neural network we use recurrent units in
of the convolution layer (receptive field). The both input section. It’s as same as SiameseCNN but
constant gradient for positive values during we use RNN instead 1d CNN layer.
backpropagation helps avoid the vanishing gradient
problem which is quite frequent in other non-linear Recurrent Neural Networks (RNN) areprimarily
activation functions like Sigmoid or Tanh. The aimed for managing data of sequential type. RNNs
formula for the function is, f(x) = max (0, x). exchange parameters around the sequence's various
position/time index/time steps, allowing instances of
Fully connected layer: different sequence lengths to be generalised
The previous layer's output would be flattened and well.RNN is typically a safer choice for position-
added to further dense layers. independent classifiers and sequential models that
handle each position differently.
Loss Layer
The loss layer is the backbone of any machine ሺ‫ݐ‬െͳሻ
learning algorithm. It is a function that takes in the ܽሺ‫ݐ‬ሻൌܾ൅‫ݓ‬ℎ ൅  ܷ‫ ݐݔ‬ሺͷሻ
ሺ‫ݐ‬ሻ
predicted value from the model and the ground truth ℎ ൌ –ƒŠ൫ܽሺ‫ݐ‬ሻ ൯ሺ͸ሻ
value, and then produces a vector or a single value ሺ‫ݐ‬ሻ
‫ ݐ ݋‬ൌ ܿ ൅ ܸℎ ሺ͹ሻ
that presents the eviation between the prediction and ሺ‫ݐ‬ሻ
‫ ݕ‬ൌ ‫ݔܽ݉ݐ݂݋ݏ‬ሺ‫ ݐ ݋‬ሻሺͺሻ
the ground truth. The goal of the training process is
to present data to the model and use optimization d.Siamese LSTM for fake news classification:
techniques that reduce the value of the loss. The LSTM unit and a basic RNN unit varies in the
notion that the LSTM contains a few virtual gates
Cross Entropy Loss and Softmax: which allows the LSTM unit to learn and retain long
The cross-entropy loss is generally used during term dependencies in data, while the RNN unit is
classification tasks where the output of the model a unable to do so. The gates are:
single number that ranges between 0 and 1 for 1. Input gate
binary classification tasks, while for multiclass 2. Forget gate
classification, the output of the model is a vector 3. Output gate
whose norm has to be 1. In cross-entropy loss, the
input to the function is interpreted as a probability 1. The internal state of the LSTM is denoted by St.
value; the probability to which class the input to the The previous internal state St-1 is combined
model’s input belongs to. If the model output is a with the next input (after linear transformation
vector, so it is possible to add the softmax function using weights) from the sequence. Instead of
to the output that normalises the vector, with each multiplication method, the use of addition
input in the vector corresponding to the probability operation enables us to escape from the problem
that the input corresponds to a certain class.The

4
Authorized licensed use limited to: Malardalen University. Downloaded on April 14,2024 at 22:02:00 UTC from IEEE Xplore. Restrictions apply.
 of vanishing gradients, as small values appear to Output gate
decrease the output. 1. The output gate is represented as:
2. Then, the combined input is passed to forget
gate which operates in a similar manner to the o = σ(bo+xtUo+ht-1Vo) (15)
input gate in the sense that it uses the sigmoid
function. This sigmoid function will operate on final output of the cell , with the tanhsquashing, can
the vector to ensure that the part of the vector expressed as
that should be forgotten are squished to values ht =tanh(s°t ) (16)
close to zero and the parts that are to be Bidirectional LSTM
remembered are moved to values closer to one.
3. This forms the internal state of the next Long Short-Term Memory (LSTM) is a tree-based
sequence. recurrent neural network used for sequential data
4. The author observed the tan squashing feature of processing with variable lengths. Bi-directional
an output sheet, the output of which is regulated LSTM makes it possible to look at individual
by an output gate. This gate specifies the values sequences from both front-to-back and back-to-front.
that are currently permissible as an output from The use of bidirectional LSTMs may superiorize
the ht cell. over other methods namely CNN, vanilla RNN and
unidirectional LSTM for fake news detection but
Input Gate drains in accuracy.
[1] Mathematical expression of tanh activation
function Architecture of Siamese LSTM
ሺܾ݃ ൅ ‫ ܷ݃ݐݔ‬൅ ℎ‫ ݐ‬െ ͳܸ݃ሺͻሻ
݃ ൌ –ƒŠ༌

Where
Ug -> weights for the input
Vg -> weights corresponding to the output of
previous cell
bg -> input bias.

Fig. 1. Sigmoid activated nodes. It will multiply the Fig 3. Siamese LSTM
output of the input gate element by element݅ ൌ
ߪሺܾ݅ ൅ ‫ ܷ݅ ݐݔ‬൅  ℎ‫ݐ‬െͳ ܸ݅ (10) Xa - HEADLINE
Xb - NEWS CONTENT
TABLE I. The output corresponding to the input S- LATER LAYERS (FC layer, output layer)
section is formulated as:
g°݅ሺͳͳሻ E.VNet the proposed architecture
->this indicates multiplication of element by element
– expressed as element eise multiplication
Forget gate and state loop

1. The forget gate or so called the state loop is

represented as:
݂ ൌ ߪሺܾ݂ ൅ ‫ ݂ܷݐݔ‬൅ ℎ‫ ݐ‬െ ͳǤ ܸ݂ሺͳʹሻ Fig 4.VNet Model
2. The combined product of the forget gate and
the element-wise product output obtained from In the VNet (the proposed architecture) we used
the previous state is represented as transfer learning model ELMo(Embedding’s from
Language Models)for embedding the news headline
st-1°f (13)
and the actual content.
3. The ultimate output from the forget gate or so
called the state loop is represented as Transfer learning
st-1°f+gi (14) Transfer learning is the process by which we take a
model that was trained for some other task, then

5
Authorized licensed use limited to: Malardalen University. Downloaded on April 14,2024 at 22:02:00 UTC from IEEE Xplore. Restrictions apply.
modify the final layers of the model and still retain The number of parameters of a machine learning
the weights of the other layers and then using this model directly relates to how easily the model can
slightly edited model to fit on other datasets. The be deployed and how quickly results can be
intuition here in computer vision is that a model that obtained. So, we analysed the number of parameters
is pre-trained on large datasets will have knowledge with the other models and found that the proposed
about different visual structures and artifacts that model contains slightly lesser number of parameters
will make it simpler for the model to fit on other than Siamese LSTM, Siamese RNN and Siamese
datasets because the pre-trained model does not have CNN, while it has approximately 600,000 more
to learn information from scratch again about visual parameters than Siamese ELMO. Despite the having
constructs and representations. For NLP tasks, the less parameters than the Siamese LSTM, Siamese
intuition for transfer learning is that a pretrained RNN and Siamese CNN, the proposed model still
model on a large corpus will already have a good performs better than competing models.
understanding about the language and other
linguistic constructs that are required for a language IV.RESULT ANALYSIS
model to do well. Unlike computer vision models,
transfer learning for NLP models started to bloom in
the recent years, but have demonstrated faster
training and better generalization.

EL Moembedding’s:
Commonly used embedding’s like word2vec or
GloVe can only assign embedding’s to individual
words. Those are word embedding’s. ELMo
embedding’s on the other hand are the output of a
function whose input is an entire sentence. Word
Fig 7. Comparison of the proposed and existing
embedding is not as effective as the nuances in a
word in various ways are captured by ELMo models
embedding.Since ELMoembedding’s are based on
entire sentences and not individual words, contextual MODEL ACCURACY PREDICITON
information about words or tokens can retrieved
here. 71.90% 89.26% SIAMESE RNN
SIAMESE CNN
SIAMESE LSTM
Layer Features SIAMESE ELMo
81.00% Vnet(proposed model)
LSTM 1,60,45,162 87.00%

CNN 1,59,17,502 77.00%

RNN 1,58,80,426
Fig 8. Comparison of Neural Network
Siamese ELMO 1, 36,00,000
ELMO-VNet 1,42,40,955

Fig 9. Model Accuracy Prediction

V.CONCLUSION AND FUTURE WORK

Identification of fake news is a challenging task in

natural language processing (NLP). Social
Fig 5.Analysing the V- Net Model networking sites' explosive growth has increased
information transparency significantly, but it has

6
Authorized licensed use limited to: Malardalen University. Downloaded on April 14,2024 at 22:02:00 UTC from IEEE Xplore. Restrictions apply.
also accelerated the spread of false information. [8] Shu, Kai, et al. "Fake news detection on social media: A data
mining perspective." ACM SIGKDD Explorations
Given the abundance of online content, automatic Newsletter19.1 (2017): 22-36.
false news detection is a necessary NLP problem for [9] Janze, Christian, and Marten Risius. "Automatic Detection of
all online content producers. In the suggested study, Fake News on Social Media Platforms." (2017).
[10] Aditi Gupta and PonnurangamKumaraguru. Credibility ranking
a study on the detection of fake news is of tweets during high impact events. In Proceedings of the 1st
described.Our paper addresses the difficulties of Workshop on Privacy and Security in Online Social Media,
PSOSM ’12, pages 2:2–2:8, New York, NY, USA, 2012. ACM.
automatically identifying fake news. We [11] W. Galuba, K. Aberer, D. Chakraborty, Z. Despotovic, and W.
systematically review the datasets using four models Kellerer. Outtweeting the Twitterers - Predicting Information
namely Siamese CNN, Siamese RNN, Siamese Cascades in Microblogs. In Usenix Workshop on Online Social
Networks, (WOSN’10), 2010.
LSTM, VNet (proposed model).The output of the [12] PonnurangamKumaraguruAnupamaAggarwal,
two layers will be merged after giving news headline AshwinRajadesingan. Phishari: Automatic realtime phishing
as one input to separate embedding layer and news detection on twitter. 7th IEEE APWG eCrime Researchers
Summit (eCRS), 2012.
material as one input to the separate embedding [13] Cook, David M., et al. "Twitter deception and influence: Issues
layer as another. The later layers that adopt unique of identity, slacktivism, and puppetry." Journal of Information
Warfare 13.1 (2014): 58-IV.
architectureis totally different from the usual fully [14] Hunt Allcott, Matthew Gentzkow. Social media and fake news
connected neural network. If we follow linear in the 2016 election. NBER, January 2017.
structure, each layer that may have important things [15] Byungkyu Kang, Tobias Hollerer, John O’Donovan. Believe it
or not? Analysing information credibility in microblogs.
to say but will be missed when we transfer the ASONAM’15 Proceedings of the 2015 IEEE/ACM
output to later layer. But theVNet catches and International Conference on Advances in Social Networks
combines the important features before the output Analysis and Mining 2015.
[16] F. Benevenuto, G. Magno, T. Rodrigues, and V. Almeida.
layer. The model gives us 81 percent accuracy, 71.9 Detecting spammers on Twitter. In CEAS, 2010.
percent, 77 percent, 89.26 percent, respectively. [17] SaptarshiGhosh, BimalViswanath, FarshadKooti, Naveen
Kumar Sharma, GautamKorlam, FabricioBenevenuto,
NiloyGanguly, and Krishna PhaniGummadi. Understanding and
REFERENCES combating link farming in the twitter social network. In
Proceedings of the 21st international conference on World Wide
[1] Nwankwo, Wilson &Ukhurebor, Kingsley. (2020). Web Forum Web, 2012.
And Social Media: A Model For Automatic Removal Of Fake [18] Pisarevskaya, Dina. "Deception Detection in News Reports in
Media Using Multilayered Neural Networks. International the Russian Language: Lexics and Discourse." Proceedings of
Journal of Scientific & Technology Research. 9. 4371- 4377. the 2017 EMNLP Workshop: Natural Language Processing
[2] Fette, N. Sadeh, and A. Tomasic, “Learning to detect phishing meets Journalism. 2017.
emails,” in Proceedings of the 16th international conference on [19] SaritaYardi, Daniel Romero, Grant Schoenebeck, and Danah
World Wide Web. ACM, 2007, pp. 649–656. Boyd. Detecting spam in a Twitter network. First Monday,
[3] Chao Yang, Robert Harkreader, Jialong Zhang, Seungwon Shin, 15(1), January 2010.
and GuofeiGu. Analysing spammers’ social networks for fun [20] Yimin Chen, Niall J.Conroy, Victoria L.Rubin. News in an
and profit: a case study of cyber Criminal ecosystem on twitter. online world: The need for an automatic crap detector. Crossref,
In Proceedings of the 21st international conference on World 2015.
Wide Web, 2012. [21] HaewoonKwak, Changhyun Lee, Hosung Park and Sue Moon.
[4] Niall J.Conroy,VictoriaL.Rubin, Yimin Chen. Automatic What is twitter, a social network or a news media? Raleigh, NC,
deception detection: Methods for finding fake news, Crossref, USA, 2010.
2015. [22] Chris Grier, Kurt Thomas, Vern Paxson, Michael Zhang.
[5] Bajaj, Samir. "“The Pope Has a New Baby!” Fake News @spam: The underground on 140 characters or less. 17th ACM
Detection Using Deep Learning." conference on Computer and communications security, 2010.
[6] PonnurangamKumaraguru, Aditi Gupta, HemankLamba. Faking
sandy : Characterizing and identifying fake images on twitter
during Hurricane Sandy. ACM, 2013.
[7] C. Shao and G. L. Ciampaglia. “Hoaxy: A platform for
Tracking Online Misinformation. 4 Mar. 2016.

7
Authorized licensed use limited to: Malardalen University. Downloaded on April 14,2024 at 22:02:00 UTC from IEEE Xplore. Restrictions apply.