Fake_News_Detection_Using_Deep_Learning_A_Systemat
Fake_News_Detection_Using_Deep_Learning_A_Systemat
This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3435497
Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2023.DOI
INDEX TERMS Classification, Deep Learning, Fake News, Misinformation, Systematic literature review
VOLUME 4, 2016 1
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3435497
false information on social media, backed with professionally Section IV describes the publicly available datasets that exist
written, long, and referenced comments that allow the reader in the fake news domain and the associated challenges. In
to more easily agree with the misinformation provided (e.g., Section V, we discuss the transfer learning strategies and
[10], [11]). Researchers aim to eliminate the increased spread open challenges in the FND context. Section VI analyses the
of misinformation by detecting the varied manners in which class imbalance problem for fake news detection. Section VII
misinformation can be spread. As such, researchers have provides a summary of the data collected in this SLR and the
resorted to the use of deep learning (DL) algorithms to detect answers to our research questions. Section VIII provides the
fake news before it spreads (e.g., [12]). This is accomplished research threats to validity and Section IX discusses the main
by collecting or creating a dataset containing both true and gaps and open issues that still exist for fake news detection.
false information within articles. Then, a pattern is deter- Lastly, Section X concludes our paper.
mined, creating a model that can predict whether a given
article contains true or false information. II. RESEARCH METHODOLOGY
There are noticeable gaps in the existing studies that were A. SEARCH STRATEGY OVERVIEW
conducted on fake news detection and that our research high- Our SLR is generated based on a set of detailed steps de-
lights. This includes (i) a lack of clear distinction between scribed in [13]. We begin by defining our research questions,
the definitions of misinformation, disinformation, and false after which we build the keywords for the search query to
information; (ii) a lack of DL-based systematic reviews on obtain the relevant papers for our study. Then, we select the
varying types of misinformation problems; (iii) a lack of gen- most relevant databases to query and establish the inclusion
eralizable DL models that allow achieving a base acceptable and exclusion criteria. Finally, we define the fields to be
detection accuracy on different datasets, which introduces the extracted from the retrieved documents.
scarce use of transfer learning in this context; and (iv) a lack
of models that deal with different levels of imbalance datasets B. RESEARCH QUESTIONS
in a fake news detection environment. The key focus of our SLR is on understanding how the DL
As technology progresses, the ability to detect misinfor- techniques have been used to address the FND problem. We
mation becomes more complicated and thus more difficult are also interested in knowing how TL has been used in this
to detect using standard machine learning (ML) techniques. field and how the class imbalance problem has been tackled.
This motivates our focus on DL techniques for the problem
• RQ1: Which deep learning algorithms have been used
of fake news detection.
for fake news detection throughout time?
In this systematic literature review (SLR), we investi-
• RQ2: Which datasets are used in the fake news detec-
gate the existing fake news detection (FND) strategies that
tion domain?
use deep learning. We then focus on the publicly available
• RQ3: How effective are deep learning methods for fake
datasets that are used in FND and their NLP approaches.
news detection?
We intended to gather information about the transfer learning
• RQ4: Which solutions are considering transfer learning
techniques applied as well as the techniques used for dealing
mechanisms (if any)?
with the class imbalance, as this will allow us to examine
• RQ5: Which solutions deal with different levels of
their effect on the detection accuracy. Our survey will help in
imbalanced datasets (if any)?
understanding the open issues and research gaps that persist
in the current research studies. As far as we know, we are the
C. SOURCE DATABASES AND SEARCH QUERY
first to provide a complete SLR that investigates the effect of
transfer learning and dealing with imbalance in the fake news We selected the following four different digital databases for
detection domain. collecting the research articles:
Key Contributions • Google Scholar (we selected the articles that appeared
The main contributions of this paper are as follows: in the first thirteen retrieved pages);
• We provide a detailed discussion of the main deep • Association for Computing Machinery (ACM) Digital
detection as well as their respective characteristics, ad- Based on the research questions established in Sec-
vantages, and disadvantages. tion II-B, we collected a set of precise concepts that can
• We study the use of transfer learning techniques and cover the topic we are studying. We, therefore, formulated
strategies for dealing with the class imbalance in this the search query as follows:
application domain. The above search statement addresses the research ques-
Paper Organization tions by focusing on the 4 key concepts in the studied topic:
This paper is organized as follows. In Section II we present “fake", “information", “detect", and “deep learning".
the research methodology. In Section III, we investigate the We searched both the title and the abstract for articles pub-
DL algorithms that are being used for detecting fake news. lished between January 2018 and December 2023 inclusive.
2 VOLUME 4, 2016
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3435497
D. INCLUSION CRITERIA
We considered the following inclusion criteria for our sys-
tematic literature review:
• Peer-reviewed journals and conference articles retrieved
from the search query defined in Figure 1.
• Articles from the Computer Science domain.
• Research articles that focus on detecting or classifying
fake news.
We applied the backward snowballing technique [14] to
gather relevant articles that might have been missed in our
search by inspecting the reference sections of the retrieved
papers. We identified two articles that were not picked up
through our search query and were added to the set of
manuscripts to analyze. FIGURE 2: PRISMA Chart of selecting and retrieving the
articles
VOLUME 4, 2016 3
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3435497
III. DEEP LEARNING ALGORITHMS USED FOR FAKE process data with a grid-like topology, such as images. They
NEWS DETECTION include traditional convolutional neural networks, residual
The thorough examination of various models and techniques networks, and dense networks.
pointed out the significant role that DL plays in different Recurrent neural networks (RNNs) are designed to handle
classification tasks including detecting fake news. Building sequential data, such as time series or language, and include
and improving such algorithms became a pressing necessity, basic recurrent neural networks and bi-directional RNNs,
especially during the COVID-19 pandemic when a large long short-term memory networks (LSTM), gated recurrent
volume of fake news and rumours were being disseminated units (GRUs) and bi-directional GRUs, and bi-directional
widely. Figure 3 demonstrates a clear increase in the use of long short-term memory networks (BiLSTM). we will refer
DL models over the years. to the model and its bi-directional version collectively as
(Bi)X, where X is the model. Graph neural networks (GNNs),
a newer type of neural network, are designed to operate
on graph-structured data, such as social networks, chemical
molecules, or protein structures.
Recently, attention-based models have gained popularity
due to their ability to focus on certain parts of the input data
selectively. They include self-attention networks and multi-
head attention networks. Hybrid models, which combine
different types of neural networks, have also become pop-
ular. For example, convolutional recurrent neural networks
combine the spatial processing capabilities of convolutional
neural networks with the temporal modeling capabilities of
recurrent neural networks. Transformer networks, like BERT,
combine self-attention mechanisms with feedforward neural
networks to process sequences of data. Figure 5 shows a
FIGURE 3: DL Models used for FND between the years of taxonomy of the various types of neural networks that are
2018 and 2023. used for FND.
Based on the data that we gathered from the surveyed
The data extracted shows that the FND task usually follows articles, it is evident that researchers extensively explored
a generic framework as is shown in Figure 4. Initially, the several DL algorithms for the detection task. Figure 6 shows
process involved acquiring or generating a dataset. The ma- the usage of different DL detection models for fake news.
jority of studies have utilized news articles that were gathered More precisely, this figure displays the percentage of papers
from openly accessible datasets. After collecting the dataset, where a particular model was used. We observe that the
pre-processing techniques were employed to prepare the data (Bi)LSTM was the most frequently included model used in
for input into a neural network. Prior investigations have 73% of papers and the CNN model was the second most used
mainly employed Word2vec and GloVe word embedding model utilized in 61% of the papers reviewed. Since multiple
methods to transform words into vectors [16]. Finally, the models may be used in the same research paper, summing
neural network model is trained and the predictions are up the percentages in Figure 6 exceeds a total of 100%. The
obtained. following sections provide a detailed discussion of the main
architecture used for FND.
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3435497
FIGURE 5: Taxonomy of the main neural network categories used for FND.
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3435497
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3435497
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3435497
and a recurrent dropout rate of 0.2. Additionally, the fully propagate information across the graph and learn representa-
connected network is designed with a batch normalization tions of the nodes and edges that capture their relationships
layer, followed by three dense layers with a ReLU activation and interactions. These learned representations can then be
function and several filters of 256, 128, and 64, respectively. used to classify the news articles as fake or real based on
The classification task is carried out using a dense layer with their similarity to other articles and the overall structure of
a softmax activation function. the graph.
Researchers focus more on testing the effects of develop- Our findings show that only 4% of the selected research
ing hybrid models that adopt LSTM in the detection process. articles adopted GNN architectures for fake news detec-
They tested the importance of (Bi)LSTMs over LSTM and tion [94], [171]–[177]. The claimed detection accuracy was
reported that the (Bi)LSTM+CNN achieved considerably incredibly low compared to the other deep-learning models
higher accuracy than when they attempted to use the LSTM on different datasets used. The highest detection accuracy
with the CNNs. They reported a detection accuracy of about obtained when using the GraphSAGE was 89.7% accuracy
99% detection accuracy when they attempted to use the without mentioning whether this was on the training or
(Bi)LSTM instead of the LSTM [97]. the testing dataset [173]. The accuracy went deeply down
When LSTM is combined with CNN, studies also reported to 61.5% when they adopted the GNN. It also recorded a
an accuracy ranging between 97.8% in [131] and 47.06% 73.12% [171] with GCN with a maximum of 88.6% [173].
in [31], with an average accuracy of 82.3%. In the case where The other variants such as SGT, GCN, and GAT had reached
LSTM is combined with a DNN architecture, we observed an an average accuracy of about 83.1%.
accuracy of 91.16% [135], while when it is combined with
BERT the accuracy achieved was 84.10% [136]. D. ATTENTION-BASED AND BERT-BASED
Bi(LSTM) is also getting popular in fake news detection ARCHITECTURES
as our survey findings show. It recorded the highest detection Another notable advancement happened in fake news de-
accuracy of 99.52% [128] and the lowest of 28% [138] with tection with the use of attention-based approaches using
an average of 75.22%. It also appeared connected to other different datasets. Our findings show that their use has been
detection architectures such as CNN and GRU. Bi(LSTM) increasing since the year of 2018 and has reached the maxi-
with CNN recorded the highest accuracy of 98.65% in [134] mum in the year of 2022. In addition, this approach appeared
and the lowest accuracy of 35.13% in [57]. The average de- in 15% of the surveyed articles mostly in the year 2022.
tection accuracy in such cases was about 77.6%. Bi(LSTM) Authors have applied it to the other detection models includ-
with GRU reached 89.8% detection accuracy [120]. ing RNNs [32], GRU [32], [76], [124], [162], [165], [178],
LSTM, and (Bi)LSTM [19], [50], [58], [78], [125], [134],
C. ARCHITECTURES BASED ON GRAPH NEURAL [142], [156], [176], [179], BERT [58], and CNN alone [50],
NETWORKS [78] or with other models [19], [50], [142]. The detection
Another popular model in fake news detection is the Graph accuracy ranged between 54% [50] and 98.65% [134].
Neural Network (GNN) and its variants such as Sequence Another deep learning model present in our surveyed
Graph Transform (SGT) [166], Graph Attention Networks works that shows cutting-edge detection is the BERT [180]
(GAT) [167], GraphSAGE [168], and Graph Convolutional model. It is a sophisticated pre-trained word-embedding
Networks(GCN) [169]. GNN is a neural network that di- model built on a transformer-encoded architecture. The find-
rectly operates on the graph structure. One of its popular ings show that 16% of the surveyed studies adopted the
applications is node classification in which every node in the BERT as a detection mechanism [44], [52], [58], [67], [75],
network has a label. This network predicts the label of the [80], [82], [84], [89], [90], [92], [100], [111], [112], [136],
node without the ground truth [170]. In FND using GNNs, [146], [158], [159], [181]–[185], [185], [186]. The findings
news articles, and related information are represented as a also show that authors started using the BERT as a detection
graph. The nodes of the graph represent the individual enti- model for fake news in 2021 which makes it still a novel
ties, such as news articles, users, or social media posts, and tool for the detection model and a future direction in the
the edges represent the relationships or interactions between fake news detection field. Our findings show that this model
them. To create the graph, the news articles are typically has reached a remarkable detection accuracy with the highest
preprocessed to extract features such as the article content, recorded accuracy of 98.5% [183] and an average accuracy of
metadata, and social media interactions. These features are around 90%. It is also clear in our findings that researchers
used to construct the nodes and edges of the graph, with the experimented with the effectiveness of applying the BERT
nodes representing the articles and the edges representing the with other models such as LSTM [136] and CNN [52] for the
relationships between articles, users, or other entities. For detection of fake news using different datasets. An example
example, edges could represent similarities between articles of using BERT in the fake news detection process, FakeBERT
or social media interactions such as retweets or mentions. has been proposed in [187] which outperforms all other
Once the graph is constructed, Graph Neural Networks are models with an accuracy of 98.9%. Figure 11 illustrates the
used to analyze the graph structure and extract useful features proposed FakeBERT.
for fake news detection. The GNNs use graph algorithms to As Figure 11 shows, this design employs three parallel
8 VOLUME 4, 2016
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3435497
blocks of 1D-CNN with 128 filters, with each block having under simple neural networks (NN; ANN, DNN, and FNN).
one convolutional layer. The first layer has a kernel size Although these models are referred to as simple detection
of 3 and 128 filters, reducing the input embedding vector techniques and were used in 7% of the total surveyed articles,
from 1000 to 998. The second layer has a kernel size of they still reached a noticeable accuracy in detecting fake
4 and 128 filters, reducing the input vector from 1000 to news. Our findings show that FFN has reached a detection
997. The third layer has a kernel size of 5 and 128 filters, accuracy of 89.8% [45] to about 95% when it was provided
decreasing the input vector from 1000 to 996. Max-pooling by solid support from a strong embedding technique [150]
layers are also included after each convolutional layer to and a lower accuracy of 83.35% [108]. On the other hand,
further reduce the dimension. A max-pooling layer with a DNN reached an accuracy of 94.68% [189] while it was less
kernel size of 5 reduces the vector to 1/5th of 996, which accurate when applying it with an LSTM by 2.8% [135]. The
is 199. After concatenating the three convolutional layers, findings also show that using a multichannel ANN [189] has
another convolution layer with a kernel size of 5 and 128 increased the detection accuracy by approximately 13% of
filters is applied. This is followed by two hidden layers with the basic ANN which was 80.9% accurate in detecting fake
384 and 128 nodes respectively. The number of trainable news [190].
parameters for each layer is also provided in the "Param In conclusion, we observe a clear growing trend in
number" column for further details. the solutions proposed using (Bi)LSTM, CNN, BERT, etc.
A recent study has conducted a thorough comparison be- throughout the years, as Figure 12 shows.
tween different deep learning models in fake news detection
using various datasets [188]. The authors studied the effect F. CHALLENGES RELATED TO DEEP LEARNING
of deploying (Bi)LSTM, CNN-RNN, C-LSTM, CNN, and METHODS FOR FAKE NEWS DETECTION
BERT in the detection of fake news. They used seven fake
Despite the promising results of deep learning methods for
news detection datasets with each model to be able to draw a
fake news detection, several challenges remain to be ad-
generalized conclusion. They figured out that the (Bi)LSTM
dressed. These include issues related to dataset quality, model
and BERT detection models achieved the best detection ac-
performance on imbalanced datasets, and the generalizability
curacies and F-scores. The authors have also concluded that
of models across different datasets. In this article, we will
BERT performs better than the (Bi)LSTM when the model
explore these challenges in more detail and discuss potential
aims at detecting fake news in different contexts from the one
solutions.
it was trained on [188].
It is important to recognize that there are several chal-
lenges when it comes to achieving effective fake news de-
E. ARCHITECTURE BASED ON FEEDFORWARD tection using deep learning methods. One major issue is the
NEURAL NETWORKS potential for overfitting, where models achieve high accuracy
Finally, other deep learning models have been used in the on the training data but perform poorly on new, unseen
fake news detection field with basic and standard feedforward data [61]. Some previous research has reported extremely
neural network (FFN) settings. Authors categorized these high accuracy results, but these were obtained by evaluating
VOLUME 4, 2016 9
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3435497
the model on the same data that was used for training. The A. MAIN DATASETS USED FOR FAKE NEWS
performance of their models achieved a high accuracy of DETECTION
99.9% [25], [54], [121], [126], [128], [133], [134], and [61]. Researchers have used several datasets in the context of fake
This raises questions about the model’s ability to generalize news detection. However, we found that only a small part of
to new data. Another challenge is the use of accuracy as the these datasets is publicly available, while a considerable per-
sole evaluation measure for imbalanced datasets, where the centage is created by the researchers and/or is not disclosed
number of fake news samples vastly outweighs the number publicly. A pie chart of the used datasets in the surveyed
of real news samples [17]–[20], [37], [58], [61]–[63], [118], studies is presented in Figure 13.
[121], [125], [133], [139]–[141]. Accuracy can be misleading
in these cases, as it can be skewed by the dominance of
the majority class. A more appropriate measure, such as
precision or recall, would provide a better understanding
of the model’s performance. Additionally, different datasets
can have varying characteristics and biases, and models that
perform well on one dataset may not generalize to other
datasets. This is noted from our findings in [44], [127],
[137], [138], and [51]. This was also proved by the thorough
experiments that were made in a recent study of cross-domain
fake news detection [188]. Finally, the quality and diversity
of the training data can greatly impact the performance of the
model [191], [192]. In some cases, models have been trained
on datasets that are not representative of the full range of fake
news content, leading to poor detection performance [192],
[193]. FIGURE 13: Datasets used in the surveyed studies
IV. DATASETS USED FOR FAKE NEWS DETECTION We observe that ISOT [194], PHEME [195], Liar [196],
In this section, we first discuss the main characteristics of and FakeNewsNet [197], with its three sub-datasets, Gossip-
the datasets used in the surveyed works. Then, we discuss Cop, PolitiFact, and BuzzFeedNews, are examples of pub-
some of the open challenges related to the datasets in this licly available fake news datasets. These are among the most
application domain. popular and frequently used datasets.
The LIAR dataset includes short statements obtained from
10 VOLUME 4, 2016
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3435497
the Politifact fact-checking website. This dataset includes a B. CHALLENGES RELATED TO THE DATASETS USED
total of 12.8 K labelled short statements. The annotation task FOR FAKE NEWS DETECTION
has been done by the Politifact site, and the statements are One of the main difficulties in the fake news detection field
classified into 6 classes: pants-fire, false, barely-true, half- is the scarcity of labelled cases [198], [199]. Even though
true, mostly-true, and true. In addition, another fake news multiple datasets with a massive amount of records exist they
dataset was collected from real-world news articles called are mostly unlabeled or have only a few records labeled.
ISOT. The real news cases were collected by crawling news Researchers have collected datasets over the last few years
articles from Reuters.com, and the fake news examples were for use with DL models in different contexts associated with
obtained from unreliable websites, which were annotated by fake news detection. Datasets are massively diverse from
the Politifact website. The PHEME dataset was collected one another due to having different research goals inside the
from Twitter based on 9 newsworthy events classified by fake news detection application domain [200]. For example,
journalists. The annotation process was conducted by jour- some datasets contain exclusively political statements, while
nalists (human annotators) and each tweet was annotated other datasets only include news articles or social media
with one of the following labels: "proven to be false", "con- posts [188].
firmed as true" or "unverified". The FakeNewsNet dataset To collect appropriate datasets to serve in fake news detec-
consists of three subdatasets, which are GossipCop, Politi- tion, we need fake articles and non-fake articles. Fake articles
Fact, and BuzzFeedNews. In total, the FakeNewsNet dataset are gathered from deceitful websites that are designed on
contains approximately 19,838 news articles labelled as ei- purpose to disseminate misinformation and fake news. The
ther "fake" or "real". The news articles in the FakeNewsNet fake news published on these websites will eventually be
dataset were annotated by a team of human annotators. The shared on social media to be read and circulated by innocent
annotators were given guidelines for identifying fake news people who do not check the news source.
and were trained to identify various characteristics of fake
It is also clear from our findings that the datasets used in
news, such as misleading headlines, fabricated content, and
fake news detection are insufficient for training models due to
misleading images. Table 1 summarizes the main character-
their characteristics, such as language features or size [201].
istics related to these datasets.
That leads us to the question of creating a dataset to serve as
TABLE 1: Main characteristics of the publicly available FND a benchmark in the detection process. However, this can be
datasets. challenging due to several reasons, some of which are:
Dataset Size Num. of Labels Type • Sources of fake and non-fake news: Identifying reliable
LIAR 12.8K 6 Political Statements sources of fake and non-fake news can be difficult,
PHEME 5800 2 Social media (tweets) especially in today’s world where there are numerous
ISOT 45k 2 News articles
sources of information and not all of them are trustwor-
thy [202]. It is crucial to ensure that the dataset contains
BuzzFeedNews 5,835k 2 News articles
a diverse range of sources to ensure that the model is
PolitiFact 12,835k 2 News articles
trained to detect fake news from a variety of sources.
GossipCop 1168k 2 News articles
• Bias in the data: Bias can be introduced in the data due
to various reasons such as the sources of the data, the
From our findings, LIAR achieved a maximum of 98.95% labelling process, or the selection criteria for the dataset.
when a Bi(LSTM) model was used for training [127]. The Bias can affect the accuracy of the model and can also
same dataset was an option for training the (Bi)GRU in [138] lead to unfair predictions [203].
which recorded a low detection accuracy of 28.12%. ISOT • Labeling issues: Labeling data for fake news detection
recorded high detection effectiveness in many cases, espe- can be challenging, as there can be discrepancies in the
cially when the trained model was a Bi(LSTM) [54] with an definition of what constitutes fake news. Human labels
accuracy of 99.95%. Still, the accuracy decreased when the may be subjective, and there may be inconsistencies
models used for training were the RNN and CNN, showing in the labelling process. Automatic labels generated
a performance of 82.5% as the lowest in [47]. PHEME using machine learning techniques can also have their
also exhibited a high performance when the CLSTM model limitations [199], [204].
was used for training, providing an accuracy of 91.88% • Bots involvement: Bots can be used to generate large
and recording a minimum accuracy of 59.2% in training volumes of fake news and spread it rapidly across the
a (Bi)GRU [124]. Lastly, FakeNewsNet sub-datasets used internet, making it difficult to detect and remove. Bots
in training different models such as CNNs [51], various can also be used to manipulate the labelling process by
RNNs [51], [111], [112], [136], [138], GNNs [172], and providing biased labels, leading to inaccuracies in the
BERT [111], [112], [136], [182]. The best detection accuracy dataset [202].
that achieved when training the GAT with a 96.42% accuracy • Rapid evolving nature of fake news: The nature of fake
while it recorded a 71.16% accuracy when it was used to train news is constantly evolving, and new techniques for
a (Bi)GRU. creating and spreading it are being developed all the
VOLUME 4, 2016 11
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3435497
time. This makes it difficult to create a comprehensive • Training models in similar domains: This transfer learn-
dataset and up-to-date [200], [205]. ing method trains models that belong to similar do-
To address these issues, it is crucial to have a well-designed mains. For instance, if there is insufficient data to com-
and diverse dataset that is regularly updated to reflect the plete task X, but task Y is similar and has adequate data,
changing nature of fake news. It is also important to have a model can be trained on task Y and then used to create
robust labelling procedures in place, using a combination a new model for task X [216].
of human and machine labels, to ensure that the dataset is • Feature extraction: Feature extraction is another trans-
unbiased and accurate. Additionally, researchers should con- fer learning approach where deep neural networks are
sider incorporating techniques such as adversarial training to trained to extract features automatically. After training
improve the robustness of the model to adversarial attacks. them on pre-existing models, the representations are
exported to new models. This technique is commonly
employed by data scientists [217].
V. STRATEGIES FOR TRANSFER LEARNING
• Utilizing pre-trained models: This approach involves
A. TRANSFER LEARNING STRATEGIES APPLIED TO
developing pre-trained models that take transfer learn-
FAKE NEWS DETECTION
ing variables into account. Companies experienced in
Numerous real-world applications have made use of the ma- model development often have access to a library
chine and deep learning techniques. These learning method- of models that can be used to create future models.
ologies assume that the input feature space and data distribu- This means that when dealing with a new problem,
tion properties are maintained across the experiments carried a pre-trained model can be selected, optimized for
out because the training data and testing data are drawn the problem at hand, and then reused to train another
from the same domain [206]. This assumption, however, may model [216].
not be accurate in some real-world machine-learning situa-
tions. In fact, in some circumstances, gathering training data The first transfer learning technique involves training mod-
can be costly and/or challenging. As a result, the research els in similar domains by using a pre-trained model from a
community has been considering the development of high- source domain that is similar or related to the target domain.
performance learners who are trained using data that could The idea is that the knowledge learned from the source
be more easily obtained from other various domains instead domain can be leveraged to improve model performance
of the deployment domain. on the target domain, even if the target domain has limited
Transfer learning is a technique used to advance a learner labelled data.
in one domain by transferring knowledge from a related Training models in similar domains typically involve the
domain. Real-world, non-technical experiences can help us following steps:
comprehend why transfer learning is feasible. Take the case 1) Selecting a source domain: The source domain should
of two individuals who wish to learn how to play the piano. be chosen based on its similarity or relevance to the
One person has no prior musical training, whereas the other target domain. Ideally, the source domain should have
plays the guitar and has a wealth of musical expertise. By similar data distribution, task, or domain characteris-
applying previously acquired musical information to the goal tics as the target domain, so the knowledge learned
of learning to play the piano, a person with a strong musical from the source domain can be effectively transferred
background will be able to learn the piano more quickly and to the target domain.
effectively [207]. One can employ knowledge from a task 2) Acquiring or creating a labelled dataset in the source
they have already mastered to help them learn a new one that domain: A labelled dataset in the source domain is
is related. needed for training the pre-trained model. This dataset
The essence and necessity of transfer learning appear when should be representative of the data in the source
there is a dearth of target training data [206]. This can be domain and should cover the task or tasks of interest.
the result of the data being rare, expensive to gather and 3) Pre-training the model on the source domain: The pre-
label, or inaccessible. The use of other existing datasets trained model is trained on the labelled dataset in the
that are related to, but not precisely the same as a given source domain. This involves training the model using
target domain of interest makes transfer learning solutions standard machine learning or deep learning techniques,
an alluring strategy since big data repositories become more such as supervised learning or unsupervised learning,
widespread. Transfer learning has been successfully used in depending on the availability of labelled data in the
many machine and deep learning applications, including text source domain.
sentiment classification [208], image classification [209]– 4) Fine-tuning or adapting the pre-trained model to the
[211], classification of human activity [212], classification of target domain: After pre-training on the source domain,
software defects [213], and classification of multi-language the pre-trained model is fine-tuned or adapted to the
text [214]. target domain. This typically involves further training
Different techniques can be utilized in transfer learning to the model using the limited labelled data available
accomplish tasks as the following [215]: in the target domain, while retaining the knowledge
12 VOLUME 4, 2016
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3435497
learned from the source domain. Fine-tuning can be 4) Training a new model on top of the extracted features:
done by updating the weights of some or all of the lay- The extracted features are then used as inputs to a new
ers of the pre-trained model, depending on the specific model, which is trained on the limited labelled data
task and data. available in the target domain. This new model, often
5) Evaluating and validating the model performance: The referred to as the target model, is trained using standard
fine-tuned model is evaluated and validated on the machine learning or deep learning techniques, such as
target domain dataset to assess its performance. This supervised learning or fine-tuning, depending on the
may involve measuring metrics such as accuracy, pre- availability of data in the target domain.
cision, recall, F1 score, or other relevant performance 5) Evaluating and validating the target model perfor-
indicators to determine the effectiveness of the transfer mance: The trained target model is evaluated and val-
learning approach. idated on the target domain dataset to assess its per-
Transfer learning by training models in similar domains formance. This may involve measuring metrics such as
can be useful when the target domain has limited labelled accuracy, precision, recall, F1 score, or other relevant
data, but related or similar domains have abundant labelled performance indicators to determine the effectiveness
data. By leveraging the knowledge learned from the related of the transfer learning approach.
source domain, the model can benefit from the additional Feature extraction in transfer learning allows leveraging
data and potentially achieve better performance on the target the knowledge learned from the source domain to extract
domain task. However, it is important to carefully consider relevant features from the data in the target domain, even
the similarity and relevance between the source and target if the target domain has limited labelled data. By using
domains to ensure that the knowledge transfer is effective and the learned features as inputs to a new model, the target
results in improved performance. model can potentially benefit from the representations or
For the second transfer learning technique, feature ex- embeddings learned from the source domain. This can help
traction is one of the common techniques used in transfer improve the performance of the target model on the target
learning, where a pre-trained model is used to extract features domain task. However, it is important to carefully consider
from data in one domain and these features are then used to the similarity and relevance between the source and target
train a new model for a different task or domain [218]. domains to ensure that the features extracted from the source
In transfer learning with feature extraction, the pre-trained domain are relevant to the target task.
model is typically a deep neural network trained on a large
For the third transfer learning technique, utilizing pre-
dataset from a source domain. This model has learned to
trained models is a common approach in transfer learn-
extract relevant features from the source domain data, which
ing where a pre-trained model, typically trained on a large
can be representations or embeddings of the input data at
dataset, is used as a starting point for training a new model
different layers of the network. These learned features are
on a smaller target dataset. The idea is that the knowledge
then used as inputs to a new model, often referred to as the
learned from the source domain can be transferred to the
target model, which is trained on the limited labelled data
target domain, even if the two domains are different, to
available in the target domain.
improve the performance of the target model [219].
The process of using feature extraction in transfer learning
typically involves the following steps: Here are some key steps involved in utilizing pre-trained
1) Selecting a pre-trained model: The pre-trained model models for transfer learning:
should be chosen based on its relevance to the target 1) Select a pre-trained model: Choose a pre-trained model
task or domain. Ideally, the pre-trained model should that is trained on a large dataset and is relevant to
have been trained on a large dataset from a source your target task. For example, suppose you are working
domain that is similar or related to the target domain, on an image classification task. In that case, you can
so that the learned features are relevant to the target choose a pre-trained Convolutional Neural Network
task. (CNN) such as VGG, ResNet, or Inception, which have
2) Removing the last layers of the pre-trained model: The been trained on large image datasets like ImageNet.
last layers of the pre-trained model, which are often 2) Remove or freeze some layers: Depending on the ar-
responsible for task-specific predictions, are removed chitecture of the pre-trained model, it may be necessary
to retain the feature extraction capability of the model. to remove or freeze some layers. For example, you can
These last layers are replaced with new layers that are remove the output layer(s) of the pre-trained model and
specific to the target task. replace them with new layers that are suitable for your
3) Extracting features from the source data: The pre- target task. Alternatively, you can freeze the weights
trained model is used to extract features from the data of some of the layers in the pre-trained model and
in the source domain. This typically involves passing only fine-tune the remaining layers during the training
the data through the layers of the pre-trained model up process.
to a certain layer and using the outputs of that layer as 3) Add new layers: Add new layers on top of the pre-
the learned features. trained model to adapt it to your target task. These new
VOLUME 4, 2016 13
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3435497
layers are typically randomly initialized and are trained works (CNNs), traditionally used for image recognition, to
using the target dataset. The output of these new layers text classification tasks, including fake news detection [22].
serves as the final prediction layer for your target task. Models like VGG16, which were previously trained on large
4) Fine-tune the model: Train the entire model, including image datasets, may be reused by replacing the last layers and
the pre-trained layers and the newly added layers, on retraining on textual data [228]. This strategy takes advantage
your target dataset. During the fine-tuning process, the of CNNs’ hierarchical feature extraction capabilities, which
weights of the pre-trained layers and the new layers are enable them to detect detailed patterns in textual data that
updated using the gradients computed from the target indicate fake news. In addition, pre-training hybrid models
dataset. Fine-tuning allows the model to learn task- on large datasets and then fine-tuning them on specific fake
specific representations while leveraging the knowl- news datasets used in fake news to exploit the strengths of
edge from the pre-trained model. both architectures.
5) Evaluate and tune: After training, evaluate the perfor- Recently, the researchers shifted the whole focus to
mance of the transferred model on your target task. You transformer-based models, particularly those like BERT,
may need to tune the hyperparameters and architecture GPT-3, and Llama [229], [230]. These models are pre-
of the transferred model to optimize its performance. trained on massive datasets using self-supervised learning
Utilizing pre-trained models can be an effective transfer techniques, which enable them to understand and generate
learning approach as it allows leveraging the knowledge human-like text. For fake news detection, these models can
learned from large datasets, reducing the need for extensive be fine-tuned on labelled datasets specific to fake news,
training data in the target domain, and potentially improving enabling them to distinguish between fake and real news with
the performance of the target model. However, it’s important high precision.
to carefully choose the pre-trained model, architecture, and Based on the collected data from the surveyed articles,
fine-tuning strategy to ensure that the transferred knowledge along with their corresponding fake news detection effective-
is relevant and beneficial for the target task. ness, the following conclusions can be drawn related to the
There are various pre-trained machine learning mod- use of transfer learning techniques in this domain:
els available in the market, such as Google’s Inception 1) CNN with AlexNet as a transfer learning technique
model [220], Microsoft’s MicrosoftML R package [221] achieved an accuracy of 93.2%. In comparison, not
and Microsoftml Python package [222], and others like applying transfer learning recorded an accuracy of
AlexNet [223], Oxford’s VGG Model [224], and Mi- 70.1% in [22].
crosoft’s ResNet [225]. In addition, some of the well- 2) In [181], pre-trained BERT as a transfer learning tech-
known pre-trained models used for NLP-related data prob- nique achieved an accuracy of 94.66%. Similarly, a
lems are Google’s word2vec Model [226], Stanford’s GloVe pre-trained BERT has also helped in the detection of
Model [227] and BERT [180]. fake news using ISOT dataset [92]. In another case
BERT is a pre-trained language model that was initially of BERT variations, RoBERTa achieved an accuracy
introduced by Google in 2018. The model is trained on a large of 92.77% and 91.7% on Politifact and Gossipcop
corpus of unlabeled text data to learn the underlying structure respectively [229] which outperform the state-of-the-
of the language. It utilizes a transformer-based architecture art, without transfer learning, techniques by achieving
that allows it to capture long-term dependencies and con- an average accuracy of 10.49% and 14.53% improve-
textual relationships between words. After pre-training, the ments on Politifact and Gossipcop, respectively.
model is fine-tuned on a specific downstream NLP task, 3) CNN with various transfer learning techniques, such
such as sentiment analysis, question answering, or named as AlexNet, ResNet50, MobileNet, DenseNet, Xcep-
entity recognition. This fine-tuning step enables the model tionNet, InceptionV3, VGG16, and VGG19, achieved
to adapt to the specific requirements of the downstream high accuracy on the EMERGENT dataset [228]. The
task. In summary, BERT is a transfer learning technique that detection accuracy was ranging between 91.22% and
leverages pre-training on unlabeled text data and fine-tuning 97.68% in [49]. VGG16 was also used as a pre-trained
on specific NLP tasks to achieve state-of-the-art performance model with freezing some layers and trained on a self-
on a variety of NLP benchmarks [180]. created dataset, achieving about 98% detection accu-
In particular, the most common transfer learning strategy racy [72].
in fake news detection is fine-tuning pre-trained models. 4) The Universal Language Model Fine-tuning transfer
Models like BERT, Llama, and GPT (Generative Pre-trained learning technique has achieved over 80% for all the
Transformer) have been pre-trained on extensive text corpora evaluation metrics (Accuracy, Precision, Recall, F1) on
and can be fine-tuned for fake news detection [92]. By PHEME dataset [56].
adjusting the weights of these models on a specific fake news
dataset, researchers can achieve high detection accuracy with B. TRANSFER LEARNING CHALLENGES FOR FAKE
relatively low computational resources. NEWS DETECTION
Other transfer learning strategies used for fake news de- Transfer learning has been used in various natural language
tection include the adaptation of Convolutional Neural Net- processing (NLP) applications, including fake news detec-
14 VOLUME 4, 2016
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3435497
tion. However, there are several challenges associated with poorly when datasets are highly imbalanced [232]. From the
applying transfer learning in this domain. standpoint of applications, the nature of the imbalance can
One initial aspect that we must highlight is that transfer be divided into two categories: data that is naturally imbal-
learning has not been extensively explored in fake news anced (e.g., credit card frauds, earthquakes, shuttle failure
detection. we consider that this is partly due to the complexity and rare diseases) or data for which it is too expensive to
of the task, which requires identifying subtle linguistic cues obtain data on the minority class for learning such as natural
and context-specific information. disasters prediction, or uncommon events prediction such as
Another challenge concerns the difficulty in finding related volcanic eruptions or tsunamis, may require historical data or
domains and publicly available datasets that can be useful for expert knowledge, which could be sparse or expensive to ob-
training the models. The success of transfer learning relies on tain [232]. This is also the case for fake news detection where
the availability of large and diverse datasets that share some the number of fake news available is much less represented
commonality with the target task. However, in the case of in the available data.
fake news detection, relevant datasets are often limited, and Several techniques have been proposed to address the
it can be challenging to find related domains that can be used issues associated with class imbalance. The three main types
for transfer learning. of techniques that can be applied are resampling techniques
The rarity of data is another significant challenge in (or data pre-processing), algorithmic level techniques, and
fake news detection. Since the detection of fake news is data post-processing techniques [233]. The solutions most
a relatively new area of research, there are limited anno- commonly used are the data pre-processing or algorithm-
tated datasets available for training and testing models. This level techniques.
scarcity of data makes it difficult to apply transfer learning In data preprocessing techniques, sampling is applied to
techniques, which rely on large amounts of labelled data for the training data to add new samples or remove existing ones.
pre-training. These techniques aim to change the training data distribution
In addition to these challenges, other issues need to be to force the learning algorithm to focus on the most relevant
addressed to apply transfer learning effectively in fake news class. This change in the training data can be accomplished
detection. For instance, the choice of pre-trained models and through over- and/or under-sampling. Over-sampling is the
their adaptation to specific tasks can significantly impact the process of adding new samples to the training data while
performance of the models. Furthermore, the transferability under-sampling is the process of removing samples. Fig-
of pre-trained models across different languages, domains, ure 14 and Figure 15 illustrate the random under-sampling
and cultures is still an active area of research. Considering and random over-sampling techniques. These techniques act
transfer learning strategies is a relevant area for further re- by randomly removing cases or adding copies of existing
search that can lead to improved solutions for FND. cases.
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3435497
93.56% respectively.
It should be noted that despite the effectiveness of class
imbalance techniques in improving the accuracy of fake news
detection models, a significant portion of the literature has
not thoroughly investigated or addressed this issue. From
our findings, only five research articles investigated the class
imbalance effect on fake news detection. This highlights the
need for further research and exploration of various imbal-
FIGURE 15: Random over-sampling
ance techniques to better understand their impact on model
performance and generalizability in the context of fake news
detection.
examples using techniques such as SMOTE (Synthetic Mi-
nority Over-sampling Technique) [234] or ADASYN (Adap-
tive Synthetic Sampling) [235]. Random over-sampling can B. CHALLENGES RELATED TO THE CLASS
help in increasing the representation of the minority class, but IMBALANCE PROBLEM IN FAKE NEW DETECTION
it may also result in overfitting or amplification of noise if not Class imbalance is a common problem in multiple applica-
done carefully. tion domains, and fake news detection is not an exception.
The second method for resolving class imbalance is to However, it has not received as much attention as it deserves
create or modify an existing algorithm. Instead of changing in the context of fake news detection, which we consider a big
the distribution of the training data, the change is applied to challenge to be addressed. The imbalance between real and
the learning and the decision process by increasing the im- fake news samples in the dataset can lead to biased classifica-
portance of the positive class. The cost-sensitive method and tion, where the model performs well on the majority class but
recognition-based approaches, kernel-based learning, such poorly on the minority class [233]. Even when considering
as support vector machine (SVM) and radial basis func- the usage of deep learning models, it was shown that the class
tion [236], are among the algorithms that have been adapted imbalance problem will still affect the performance of the
to address the class imbalance problem. Typically, specially models [238].
developed algorithms for dealing with the class imbalance One of the challenges related to the class imbalance prob-
issue will work very well for a specific domain for which they lem in fake news detection is the issue of using adequate
were thought. However, they will fail under other domains performance assessment metrics to evaluate the model’s
and they require a thorough understanding of the algorithm performance. Traditional metrics such as accuracy can be
to implement the modifications [233]. misleading, as the model may perform well on the majority
Our findings show that the use of various imbalance class but miss out on correctly identifying the minority class.
techniques, such as oversampling and downsampling, has This issue emphasizes the need for specialized metrics such
shown promising results in improving the performance of as F1 score, precision, and recall [239].
different classifiers, including RNN variations, CNN, and In the FND domain, there is a lack of systematic studies
hybrid models like CNN+LSTM. The results indicate that that evaluate the impact of known techniques for dealing
oversampling has been effective in improving the accuracy with class imbalance. Techniques such as oversampling,
of LSTM and CNN models in [21], achieving accuracies of undersampling, and ensemble methods have been widely
95.51% and 98.96%, respectively. Similarly, oversampling used in many other domains. However, their effectiveness in
has also been beneficial for BERT, achieving an accuracy of fake news detection remains understudied. Therefore, more
94.66% in [181]. research is needed to explore the effectiveness of these tech-
Moreover, the use of SMOTE oversampling has demon- niques in the FND domain. An important challenge with the
strated effectiveness in dealing with class imbalance. For application of these techniques for fake news detection is
instance, in [237], SMOTE was used to improve the perfor- related to the generation of fake news texts. In this case, it
mance of a DNN model, achieving an accuracy of 98% on is necessary to generate complete texts that look like real
the Politifact dataset. news, but it is also necessary to generate texts that correspond
Additionally, another study [179] utilized the focal loss to fake news. This leads to another challenge connected
function to prevent classification bias towards the majority to the need to carefully craft the synthetic text generation
class, which significantly improved the performance of their so that it corresponds to either fake news or real news. In
models on imbalanced datasets. particular, the generated fake news articles should be realistic
Downsampling has shown effectiveness in improving the and representative of the actual fake news articles to ensure
training accuracy of hybrid models like CNN+LSTM on the effectiveness of the model.
PHEME and FN-COV datasets in [35], achieving accu- The context is also a challenge when considering the gen-
racy rates of 91.88% and 98.62% respectively. Addition- eration of fake news. Since fake news is often generated in
ally, downsampling has improved the accuracy of CNN and response to specific events or situations, it can be difficult to
LSTM models in [53], achieving accuracies of 92.38% and apply generic techniques for dealing with a class imbalance
16 VOLUME 4, 2016
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3435497
that does not consider the specific context in which the fake websites intended to disseminate misinformation. This fake
news was generated. news will end up being shared on social media platforms
Lastly, special-purpose algorithms that can deal with the by its creator. This fake news will also be shared by other
class imbalance problem have not been explored or evaluated individuals unintentionally without checking the news source
for FND. These algorithms include cost-sensitive learning, or by other malicious users and bots.
manipulating the loss functions, or building ensembles that Our findings show that Liar, ISOT, PHEM, and Fake-
are specially developed to address the class imbalance prob- NewsNet (with their three variations) are the most popular
lem [233]. These techniques have shown promising results datasets being used in fake news detection. These six datasets
in multiple domains, and their effectiveness in FND requires have been used in about 80% of the surveyed articles. We
further investigation. also noticed that researchers frequently attempted to create
In conclusion, addressing the class imbalance problem in their own dataset to reach the required size and the domain
fake news detection is crucial for developing accurate and which is obvious in about 45% of the surveyed studies.
reliable models. Still, not much research has been done to Other researchers combined two or more datasets to have an
address this problem. Researchers and practitioners need acceptable-sized dataset.
to pay more attention to this problem and explore various It is also worth mentioning that selecting a proper dataset
techniques to overcome it. This is a possible area where is a crucial task in fake news detection since it will impact the
future researchers should focus on that may lead to improved detection effectiveness. It is noticeable from our findings that
solutions for FND. applying the same detection model in different datasets has
an enormous difference in the detection accuracy [28], [31],
VII. ANSWERS TO RESEARCH QUESTIONS [42]–[44], [47], [51], [123], [127], [137], [138], [165], [172],
In this section, we attempt to answer the research questions [174], [188], [190].
presented in Section II-B based on our findings. The detailed RQ3: How effective are deep learning methods for fake
answers are described below. news detection?
RQ1: Which algorithms are used for fake news detec-
Researchers studied various DL algorithms in the detection
tion throughout time?
and classification of fake news as we mentioned previously.
Given our findings, deep learning models are considered
These algorithms include CNN, RNN (with it is variations),
effective models in fake news detection. There is a notable
GNN, BERT and Attention-based mechanisms, and hybrid
increase in the number of articles that address the different
approaches. The detection effectiveness of these algorithms
models and architectures for this task. We also noticed that
is affected by the datasets used and by the use of a combina-
the research focus shifted towards deep learning models for
tion of different architectures for the detection.
FND during the global COVID-19 Pandemic in 2021 which
CNN and Bi(LSTM) have been the most used detection
forms about 92% of the research effort that was conducted on
models and they achieved the highest detection accuracy
FND. The remaining 8% of the FND research was conducted
when compared against the other approaches. RNNs, in-
before this year.
cluding their variations such as LSTM/(Bi)LSTM and GRU,
Our findings also show that fake news can be detected
are utilized with considerable effectiveness at 70%. Their
by CNNs, RNNs, GRUs, LSTMs, and BERTs models in
ability to maintain information over sequences allows them
many variations and with different architectures. We noticed
to understand context better, which is essential for identifying
that LSTM/(Bi)LSTM were the models that appeared more
fake news. CNNs on the other hand have proven to be
frequently in the surveyed articles. The detection was also
effective for fake news detection tasks, appearing in 61% of
examined using hybrid models which increased the detection
the research articles we surveyed. BERT and hybrid detection
effectiveness at some points. It is also noticeable that using
models have also made a noticeable detection effectiveness
the BERT model in the detection of fake news exhibits a huge
appearing in about 47% of the surveyed articles. Feedforward
positive impact on the detection effectiveness.
Neural Networks and Graph Neural Networks were also used
RQ2: Which datasets are used in the fake news detec-
in the detection process even though not in many studies.
tion domain?
The most difficult part of detecting fake news is the ab- It is worth mentioning that in one research article, many
sence of a labelled dataset with trustworthy ground truth deep learning models were developed to draw comprehensive
labels with an accepted size [197]. For several usages in conclusions. Hence, the total percentage of all the models that
DL, researchers attempted to collect datasets over the last appeared in the surveyed articles is more than 100%.
few years. The collected datasets are massively varied from The detailed effectiveness of DL detection models in the
one another due to the purpose of the study. For instance, fake news field is in Section III.
some of these datasets are political and consist of political RQ4: Which solutions are considering transfer learn-
statements as is the case in PolitiFact. Other datasets are built ing mechanisms (if any)?
with news articles collected in a specific time frame, while Transfer learning is the process of exploiting what has
other datasets include social media posts such as Twitter. been learned in one task to improve the generalization in
Moreover, fake news is frequently collected from duplicitous another task [206]. The goal of transfer learning is to improve
VOLUME 4, 2016 17
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3435497
learning in the target task by leveraging knowledge from the to well-classified examples and focusing more on hard-to-
source task. classify instances [179]. This approach helps to prevent the
Transfer learning is not applied in many fake news detec- model from becoming biased towards the majority class and
tion studies as our findings show. There are only seven re- ensures that the minority class instances are given appropriate
search articles that examined the effect of transfer learning on attention during training.
detection accuracy [22], [72], [92], [99], [181], [185], [229]. Handling the imbalanced dataset achieved a better accu-
However, utilizing transfer learning strategies increased the racy result compared to the baseline cases that do not deal
detection accuracy. The transfer learning that was utilized in with the class imbalance. Thus, this is a relevant area for
the FND domain may be categorized under fine-tuning pre- further research that can lead to improved solutions for FND.
trained models, using CNN-based architectures, employing
pre-trained hybrid models, and leveraging transformer-based VIII. THREATS TO VALIDITY
models. The highest improvement presented by utilizing SLRs are prone to several threats to validity that may lead to
transfer learning was by reaching an accuracy of 93.2% when a bias in the review outcomes. These threats are publication
applying the Alexnet pre-trained model which represents an bias and errors in data collection, study exclusion, and data
improvement of 23.1% compared to the baseline case which extraction. Regarding publication bias, studies with positive
is done without applying transfer learning. results are more expected to be selected over negative studies.
It is worth mentioning that applying the same detection This issue is alleviated by attempting to determine whether
model on different datasets recorded enormous differences in the studies discuss their results and limitations. Moreover, the
the detection accuracy [28], [29], [31], [42], [44], [47], [51], sole purpose of this SLR is to report the effectiveness of DL
[123], [127], [137], [138], [165], [172], [174], [188], [190]. models rather than present new results. In addition, there is
This issue might be tackled by including a transfer learning no motivation from our SLR to select studies reporting only
approach so the detection model can report an approximate positive results.
accuracy. Regarding filtering out studies based on the search criteria,
RQ5: Which solutions deal with different levels of an we aimed to have a broad search query as we mentioned in
imbalanced dataset? Section II-C to alleviate this threat. We could also expand the
A dataset with a skewed class distribution where the end- survey date range to contain the studies that were published
user preferences are biased towards the least represented before 2018. However, fake news became more popular from
class(es) suffers from a class imbalance problem. A model 2018 onward, and we aimed to provide an updated review
learned under these conditions will focus on the majority of the most recent trends in this application domain. This
class and will not learn correctly the minority and important motivation is supported by Figure 3 which demonstrates the
classes [233]. remarkable increase in fake news detection publications over
Most of the available fake news datasets are imbalanced. time.
From the articles we surveyed, only 7 papers specifically
Regarding the issue of incorrectly excluded articles and
treated class imbalance and studied its effect on fake news
extracting the data, we alleviated this issue by asking another
detection by utilizing various strategies to handle this is-
researcher to review some random studies. There is no rule
sue. These strategies were: random and advance oversam-
for determining the number of articles for the random check
pling in four articles, random undersampling in two articles,
task, but about half of the surveyed articles were selected for
and utilizing a different loss function in one article. The
this special check.
oversampling has been deployed by increasing the number
of instances in the minority class to match the majority
IX. MAIN GAPS AND OPEN ISSUES
class which improved the detection effectiveness [21], [32],
[181]. In addition, advanced oversampling techniques such as From our investigation, we gathered a list of the main gaps
the Synthetic Minority Over-sampling Technique (SMOTE), and open challenges that still deserve the attention of the
generate synthetic examples of the minority class by inter- research community for the fake news detection problem. We
polating between existing instances rather than duplicating must highlight that this is a challenging task, involving sev-
existing ones. SMOTE helped the trained model get about eral difficulties which we describe to allow future researchers
95% detection accuracy with a noticeable improvement com- to focus on the most important open issues.
pared to the baseline case without treating the class imbal- • Lack of labelled data: One of the major challenges in
ance [237]. This helped to mitigate the risk of overfitting and training deep learning models for fake news detection
enhanced the model’s generalization ability. is the limited availability of labelled data [199], [204].
Another strategy presented to balance the dataset was the Fake news datasets are often small, and obtaining ac-
random undersampling which involves reducing the number curate and comprehensive annotations for training can
of instances in the majority class to match the minority be challenging. This can impact the performance and
class [35], [53]. generalization of deep learning models, as they heavily
Finally, the focal loss function is designed to address rely on large amounts of labelled data for effective
the class imbalance by down-weighting the loss assigned training.
18 VOLUME 4, 2016
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3435497
• Potentially biased datasets: Another issue in fake news verse sources of misinformation, and rapid information
detection is the potential bias in the datasets used for spread, can impact the performance and reliability of
training and evaluation [203]. Fake news datasets may deep learning models in practical scenarios [200], [205].
contain inherent biases, such as political or cultural
biases, that can affect the performance and fairness of X. CONCLUSION
deep learning models. It is essential to carefully curate The increasing volume of people using communication plat-
and preprocess datasets to mitigate these biases and forms has opened the door for the spread of fake news. Fake
ensure the reliability and generalizability of the models. news can influence readers in many aspects, and it is crucial
• Lack of benchmarks: There is a lack of standardized to understand this phenomenon and study mechanisms that
benchmarks for evaluating the performance of deep allow its early detection. Deep learning has shown its poten-
learning models in fake news detection. The absence of tial in various tasks, including natural language processing,
benchmark datasets, evaluation metrics, and protocols and our systematic literature review highlights its effective-
makes it challenging to compare the performance of ness in fake news detection.
different models and assess their effectiveness [188], From our findings, the main categories of algorithms used
[240]. The development of standardized benchmarks for FND are CNN, RNN, GNN, Attention-based mecha-
can facilitate fair and rigorous comparisons and foster nisms, and BERT. Among these, the most frequently used are
advancements in the field. RNN-based models, which include the Bi(LSTM). We also
• Transfer learning solutions not sufficiently explored: found that Liar, ISOT, PHEME, and FakeNewsNet are the
Transfer learning, which leverages pre-trained models publicly available datasets most frequently used in fake news
for feature extraction or model initialization, has shown detection. These datasets are a central aspect because select-
promise in improving the performance of deep learn- ing a proper dataset is crucial. In effect, the data selection
ing models for various tasks [241], [242]. However, in will have an important impact on the detection effectiveness.
the context of fake news detection, the exploration of Finally, we found that transfer learning and the class
transfer learning solutions is still limited. There is a need imbalance problem are not widely explored in fake news
to further investigate and optimize transfer learning ap- detection studies, even though these techniques have shown
proaches for fake news detection to leverage knowledge promising results in increasing detection accuracy in many
from related tasks and domains. fields. Overall, our systematic review highlights the potential
• Class imbalance not adequately addressed: Class of deep learning in fake news detection and identifies impor-
imbalance, where the number of samples in different tant areas for future research. We also provide a comprehen-
classes is significantly imbalanced, is a common issue sive list of the main gaps and open issues in this domain to
in fake news detection. Deep learning models trained on guide the next steps of research in this area.
imbalanced datasets may result in biased and inaccurate
predictions, as they tend to be biased towards the ma-
jority class [243], [244]. Although some studies have
explored imbalance techniques such as oversampling
or undersampling, the effectiveness of these techniques
in deep learning for fake news detection needs further
investigation.
• Limited understanding of fake news dynamics: Despite
MOHAMMAD Q. ALNABHAN holds a Bache-
extensive research on fake news, there is still a limited lor’s and Master’s degree in Computer Science
understanding of the complex dynamics and mecha- with a specialization in data mining. Currently
nisms underlying the spread and impact of misinforma- pursuing a Ph.D. at the University of Ottawa,
tion [245]. Deep learning models for fake news detec- Canada. His major field of study is focused on
deep learning for security matters and fake news
tion may be limited by the lack of a comprehensive un- detection on social media.
derstanding of how fake news is created, disseminated, He brings a wealth of expertise to his academic
and received, which can impact the models’ accuracy pursuits through a diverse range of work experi-
and effectiveness. Further research is needed to better ences. Mohammad has been a dedicated part-time
professor in the Computer Science department at the University of Ottawa
understand the underlying dynamics of fake news and since 2022, where he imparts his knowledge and passion for the subject to
inform the development of more effective solutions. students. In addition to his academic role, he has proven his skills as a full-
• Real-world applicability: While deep learning models stack Programmer and a Microsoft Certified Trainer Specialist, showcasing
for fake news detection show promising results in con- his proficiency in industry-standard technologies.
trolled research settings, their real-world applicability,
and effectiveness in detecting fake news in diverse and
dynamic environments, such as social media or online
news platforms, is still a challenge. Real-world fac-
tors, such as varying levels of information quality, di-
VOLUME 4, 2016 19
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3435497
PAULA BRANCO Paula Branco received the [18] A. Abdullah, M. Awan, M. Shehzad, and M. Ashraf, “Fake news classi-
Ph.D. degree in Computer Science from the Fac- fication bimodal using convolutional neural network and long short-term
ulty of Sciences of the University of Porto, Portu- memory,” Int. J. Emerg. Technol. Learn, vol. 11, pp. 209–212, 2020.
gal, in 2018. She joined the School of Electrical [19] A. Abedalla, A. Al-Sadi, and M. Abdullah, “A closer look at fake news
Engineering and Computer Science, University of detection: A deep learning perspective,” in Proceedings of the 2019 3rd
Ottawa, as an Assistant Professor in January 2020. international conference on advances in artificial intelligence, 2019, pp.
Her research interests are focused on Artificial 24–28.
Intelligence, Machine Learning, KDD, and Data [20] M. Al-Sarem, A. Alsaeedi, F. Saeed, W. Boulila, and O. AmeerBakhsh,
Mining, and in particular, imbalanced domain “A novel hybrid deep learning model for detecting covid-19-related
rumors on social media based on lstm and concatenated parallel cnns,”
problems, rare extreme values forecast, health and
Applied Sciences, vol. 11, no. 17, p. 7940, 2021.
cybersecurity applications, and privacy.
[21] M. N. Alenezi and Z. M. Alqenaei, “Machine learning in detecting covid-
19 misinformation on twitter,” Future Internet, vol. 13, no. 10, p. 244,
2021.
REFERENCES [22] N. M. AlShariah, A. Khader, and J. Saudagar, “Detecting fake images on
[1] H. Allcott and M. Gentzkow, “Social media and fake news in the 2016 social media using machine learning,” International Journal of Advanced
election,” Journal of economic perspectives, vol. 31, no. 2, pp. 211–236, Computer Science and Applications, vol. 10, no. 12, pp. 170–176, 2019.
2017. [23] M. Z. Asghar, A. Habib, A. Habib, A. Khan, R. Ali, and A. Khattak, “Ex-
[2] M. R. Islam, M. A. Kabir, A. Ahmed, A. R. M. Kamal, H. Wang, and ploring deep neural networks for rumor detection,” Journal of Ambient
A. Ulhaq, “Depression detection from social network data using machine Intelligence and Humanized Computing, vol. 12, pp. 4315–4333, 2021.
learning techniques,” Health information science and systems, vol. 6, pp. [24] M. C. Buzea, S. Trausan-Matu, and T. Rebedea, “Automatic fake news
1–12, 2018. detection for romanian online news,” Information, vol. 13, no. 3, p. 151,
[3] H. Gao and H. Liu, “Data analysis on location-based social networks,” 2022.
Mobile social networking: an innovative approach, pp. 165–194, 2014. [25] M. K. Elhadad, K. F. Li, and F. Gebali, “An ensemble deep learn-
[4] K. Sharma, F. Qian, H. Jiang, N. Ruchansky, M. Zhang, and Y. Liu, ing technique to detect covid-19 misleading information,” in Ad-
“Combating fake news: A survey on identification and mitigation tech- vances in Networked-Based Information Systems: The 23rd International
niques,” ACM Transactions on Intelligent Systems and Technology Conference on Network-Based Information Systems (NBiS-2020) 23.
(TIST), vol. 10, no. 3, pp. 1–42, 2019. Springer, 2021, pp. 163–175.
[5] L. Wu, J. Li, X. Hu, and H. Liu, “Gleaning wisdom from the past: Early [26] K. M. Fouad, S. F. Sabbeh, and W. Medhat, “Arabic fake news detection
detection of emerging rumors in social media,” in Proceedings of the using deep learning,” CMC-Comput. Mater. Contin, vol. 71, pp. 3647–
2017 SIAM international conference on data mining. SIAM, 2017, pp. 3665, 2022.
99–107. [27] M. Z. H. George, N. Hossain, M. R. Bhuiyan, A. K. M. Masum, and
[6] L. Wu, F. Morstatter, K. M. Carley, and H. Liu, “Misinformation in S. Abujar, “Bangla fake news detection based on multichannel combined
social media: definition, manipulation, and detection,” ACM SIGKDD cnn-lstm,” in 2021 12th International Conference on Computing Com-
explorations newsletter, vol. 21, no. 2, pp. 80–90, 2019. munication and Networking Technologies (ICCCNT). IEEE, 2021, pp.
1–5.
[7] J. Ma, W. Gao, P. Mitra, S. Kwon, B. J. Jansen, K.-F. Wong, and M. Cha,
“Detecting rumors from microblogs with recurrent neural networks,” [28] M. H. Goldani, R. Safabakhsh, and S. Momtazi, “Convolutional neural
in Proceedings of the 25th International Joint Conference on Artificial network with margin loss for fake news detection,” Information Process-
Intelligence (IJCAI 2016). AAAI Press, 2016, pp. 3818–3824. ing & Management, vol. 58, no. 1, p. 102418, 2021.
[29] S. Gonwirat, A. Choompol, and N. Wichapa, “A combined deep learning
[8] S. K. Bharti, R. Pradhan, K. S. Babu, and S. K. Jena, “Sarcasm analysis
model based on the ideal distance weighting method for fake news
on twitter data using machine learning approaches,” Trends in Social
detection,” International Journal of Data and Network Science, vol. 6,
Network Analysis: Information Propagation, User Behavior Modeling,
no. 2, pp. 347–354, 2022.
Forecasting, and Vulnerability Assessment, pp. 51–76, 2017.
[30] O. A. Hanshal, O. N. Ucan, and Y. K. Sanjalawe, “Hybrid deep learning
[9] S. Helmstetter and H. Paulheim, “Weakly supervised learning for fake
model for automatic fake news detection,” Applied Nanoscience, pp. 1–
news detection on twitter,” in 2018 IEEE/ACM International Conference
11, 2022.
on Advances in Social Networks Analysis and Mining (ASONAM).
IEEE, 2018, pp. 274–277. [31] Y.-F. Huang and P.-H. Chen, “Fake news detection using an ensemble
learning model based on self-adaptive harmony search algorithms,” Ex-
[10] S. Kumar and N. Shah, “False information on web and social media: A
pert Systems with Applications, vol. 159, p. 113584, 2020.
survey,” arXiv preprint arXiv:1804.08559, 2018.
[32] V.-I. Ilie, C.-O. Truică, E.-S. Apostol, and A. Paschke, “Context-aware
[11] K. Shu, L. Cui, S. Wang, D. Lee, and H. Liu, “defend: Explainable fake misinformation detection: A benchmark of deep learning architectures
news detection,” in Proceedings of the 25th ACM SIGKDD international using word embeddings,” IEEE Access, vol. 9, pp. 162 122–162 146,
conference on knowledge discovery & data mining, 2019, pp. 395–405. 2021.
[12] R. K. Kaliyar, A. Goswami, P. Narang, and S. Sinha, “Fndnet–a deep [33] K. Ivancová, M. Sarnovskỳ, and V. Maslej-Krcšñáková, “Fake news
convolutional neural network for fake news detection,” Cognitive Sys- detection in slovak language using deep learning techniques,” in 2021
tems Research, vol. 61, pp. 32–44, 2020. IEEE 19th World Symposium on Applied Machine Intelligence and
[13] S. Keele et al., “Guidelines for performing systematic literature reviews Informatics (SAMI). IEEE, 2021, pp. 000 255–000 260.
in software engineering,” 2007. [34] Y. Ji, “Fake news detection based on a bi-directional lstm with cnn,” in
[14] S. Jalali and C. Wohlin, “Systematic literature studies: database searches Computing and Data Science: Third International Conference, CONF-
vs. backward snowballing,” in Proceedings of the ACM-IEEE interna- CDS 2021, Virtual Event, August 12-17, 2021, Proceedings. Springer,
tional symposium on Empirical software engineering and measurement, 2022, pp. 36–44.
2012, pp. 29–38. [35] R. K. Kaliyar, A. Goswami, and P. Narang, “A hybrid model for effective
[15] J. Babineau, “Product review: Covidence (systematic review soft- fake news detection with a novel covid-19 dataset.” in ICAART (2), 2021,
ware),” Journal of the Canadian Health Libraries Association/Journal de pp. 1066–1072.
l’Association des bibliothèques de la santé du Canada, vol. 35, no. 2, pp. [36] R. K. Kaliyar, A. Mohnot, R. Raghhul, V. Prathyushaa, A. Goswami,
68–71, 2014. N. Singh, and P. Dash, “Multideepfake: Improving fake news detection
[16] S. Girgis, E. Amer, and M. Gadallah, “Deep learning algorithms for with a deep convolutional neural network using a multimodal dataset,”
detecting fake news in online text,” in 2018 13th international conference in Advanced Computing: 10th International Conference, IACC 2020,
on computer engineering and systems (ICCES). IEEE, 2018, pp. 93–97. Panaji, Goa, India, December 5–6, 2020, Revised Selected Papers, Part I
[17] Q. Abbas, M. U. Zeshan, and M. Asif, “A cnn-rnn based fake news 10. Springer, 2021, pp. 267–279.
detection model using deep learning,” in 2022 International Seminar on [37] R. K. Kaliyar, R. Singh, S. N. Laya, M. S. Sudharshan, A. Goswami,
Computer Science and Engineering Technology (SCSET). IEEE, 2022, and D. Garg, “Rumeval2020-an effective approach for rumour detection
pp. 40–45. with a deep hybrid c-lstm model,” in Advanced Computing: 10th Interna-
20 VOLUME 4, 2016
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3435497
tional Conference, IACC 2020, Panaji, Goa, India, December 5–6, 2020, [58] A. Wani, I. Joshi, S. Khandve, V. Wagh, and R. Joshi, “Evaluating deep
Revised Selected Papers, Part I 10. Springer, 2021, pp. 300–312. learning approaches for covid19 fake news detection,” in Combating
[38] A. Zubiaga, A. Aker, K. Bontcheva, M. Liakata, and R. Procter, “De- Online Hostile Posts in Regional Languages during Emergency Situation:
tection and resolution of rumours in social media: A survey,” ACM First International Workshop, CONSTRAINT 2021, Collocated with
Computing Surveys (CSUR), vol. 51, no. 2, pp. 1–36, 2018. AAAI 2021, Virtual Event, February 8, 2021, Revised Selected Papers
[39] V. L. Rubin, Y. Chen, and N. K. Conroy, “Deception detection for news: 1. Springer, 2021, pp. 153–163.
three types of fakes,” Proceedings of the Association for Information [59] Z. Wang, Z. Yin, and Y. A. Argyris, “Detecting medical misinformation
Science and Technology, vol. 52, no. 1, pp. 1–4, 2015. on social media using multimodal deep learning,” IEEE journal of
[40] J. Brummette, M. DiStaso, M. Vafeiadis, and M. Messner, “Read all biomedical and health informatics, vol. 25, no. 6, pp. 2193–2203, 2020.
about it: The politicization of “fake news” on twitter,” Journalism & Mass [60] F. Xing and C. Guo, “Mining semantic information in rumor detection
Communication Quarterly, vol. 95, no. 2, pp. 497–517, 2018. via a deep visual perception based recurrent neural networks,” in 2019
[41] A. Marlatt, “Records suggest 2020 election conspir- IEEE International Congress on Big Data (BigDataCongress). IEEE,
acy involved 80m people,” https://www.satirewire.com/ 2019, pp. 17–23.
claim-anti-trump-conspiracy-involved-80-million-people/, 2020, [61] A. Zervopoulos, A. G. Alvanou, K. Bezas, A. Papamichail,
online, Accessed: 2022-11-11. M. Maragoudakis, and K. Kermanidis, “Deep learning for fake
[42] A. J. Keya, S. Afridi, A. S. Maria, S. S. Pinki, J. Ghosh, and M. Mridha, news detection on twitter regarding the 2019 hong kong protests,” Neural
“Fake news detection based on deep learning,” in 2021 International Computing and Applications, vol. 34, no. 2, pp. 969–982, 2022.
Conference on Science & Contemporary Technologies (ICSCT). IEEE, [62] Y. Wang, F. Ma, Z. Jin, Y. Yuan, G. Xun, K. Jha, L. Su, and J. Gao, “Eann:
2021, pp. 1–6. Event adversarial neural networks for multi-modal fake news detection,”
[43] P. M. Konkobo, R. Zhang, S. Huang, T. T. Minoungou, J. A. Ouedraogo, in Proceedings of the 24th acm sigkdd international conference on
and L. Li, “A deep learning model for early detection of fake news on knowledge discovery & data mining, 2018, pp. 849–857.
social media,” in 2020 7th International Conference on Behavioural and [63] R. K. Kaliyar, “Fake news detection using a deep neural network,” in
Social Computing (BESC). IEEE, 2020, pp. 1–6. 2018 4th International Conference on Computing Communication and
[44] R. Kozik, S. Kula, M. Choraś, and M. Woźniak, “Technical solution to Automation (ICCCA). IEEE, 2018, pp. 1–7.
counter potential crime: Text analysis to detect fake news and disinfor- [64] O. Ajao, D. Bhowmik, and S. Zargari, “Fake news identification on
mation,” Journal of Computational Science, vol. 60, p. 101576, 2022. twitter with hybrid cnn and rnn models,” in Proceedings of the 9th
[45] V. M. Krešňáková, M. Sarnovskỳ, and P. Butka, “Deep learning methods international conference on social media and society, 2018, pp. 226–230.
for fake news detection,” in 2019 IEEE 19th International Symposium [65] L. Wu, Y. Rao, H. Yu, Y. Wang, and A. Nazir, “False information
on Computational Intelligence and Informatics and 7th IEEE Interna- detection on social media via a hybrid deep model,” in Social Informatics:
tional Conference on Recent Achievements in Mechatronics, Automa- 10th International Conference, SocInfo 2018, St. Petersburg, Russia,
tion, Computer Sciences and Robotics (CINTI-MACRo). IEEE, 2019, September 25-28, 2018, Proceedings, Part II 10. Springer, 2018, pp.
pp. 000 143–000 148. 323–333.
[46] E. Masciari, V. Moscato, A. Picariello, and G. Sperlí, “Detecting fake [66] K. Popat, S. Mukherjee, A. Yates, and G. Weikum, “Declare: Debunking
news by image analysis,” in Proceedings of the 24th symposium on fake news and false claims using evidence-aware deep learning,” arXiv
international database engineering & Applications, 2020, pp. 1–5. preprint arXiv:1809.06416, 2018.
[47] J. A. Nasir, O. S. Khan, and I. Varlamis, “Fake news detection: A [67] S. Alyoubi, M. Kalkatawi, and F. Abukhodair, “The detection of fake
hybrid cnn-rnn based deep learning approach,” International Journal of news in arabic tweets using deep learning,” Applied Sciences, vol. 13,
Information Management Data Insights, vol. 1, no. 1, p. 100007, 2021. no. 14, p. 8209, 2023.
[48] A. Priya and A. Kumar, “Deep ensemble approach for covid-19 fake news [68] G. Güler and S. GÜNDÜZ, “Deep learning based fake news detection
detection from social media,” in 2021 8th International Conference on on social media,” International Journal of Information Security Science,
Signal Processing and Integrated Networks (SPIN). IEEE, 2021, pp. vol. 12, no. 2, pp. 1–21, 2023.
396–401. [69] Y. Doke, P. Dongare, M. Gaikwad, M. Gaikwad, and V. Marathe, “Deep
[49] C. Raj and P. Meel, “Convnet frameworks for multi-modal fake news fake detection through deep learning,” International Journal for Research
detection,” Applied Intelligence, pp. 1–17, 2021. in Applied Science & Engineering Technology (IJRA), vol. 11, no. 5, pp.
[50] S. Ramya and R. Eswari, “Attention-based deep learning models for 861–866, 2023.
detection of fake news in social networks,” International Journal of [70] Y. Lu and H. Ye, “Detection method of fake news spread in social net-
Cognitive Informatics and Natural Intelligence (IJCINI), vol. 15, no. 4, work based on deep learning,” in International Conference on Advanced
pp. 1–25, 2021. Hybrid Information Processing. Springer, 2022, pp. 473–488.
[51] H. Saleh, A. Alharbi, and S. H. Alsamhi, “Opcnn-fake: Optimized [71] F. Mira, “Deep learning technique for recognition of deep fake videos,”
convolutional neural network for fake news detection,” IEEE Access, in 2023 IEEE IAS Global Conference on Emerging Technologies (Glob-
vol. 9, pp. 129 471–129 489, 2021. ConET). IEEE, 2023, pp. 1–4.
[52] M. Samadi, M. Mousavian, and S. Momtazi, “Persian fake news detec- [72] C. Mallick, S. Mishra, and M. R. Senapati, “A cooperative deep learning
tion: Neural representation and classification at word and text levels,” model for fake news detection in online social networks,” Journal of
Transactions on Asian and Low-Resource Language Information Pro- Ambient Intelligence and Humanized Computing, vol. 14, no. 4, pp.
cessing, vol. 21, no. 1, pp. 1–11, 2021. 4451–4460, 2023.
[53] M. Sarnovskỳ, V. Maslej-Krešňáková, and K. Ivancová, “Fake news [73] M. Madani, H. Motameni, and R. Roshani, “Fake news detection using
detection related to the covid-19 in slovak language using deep learning feature extraction, natural language processing, curriculum learning,
methods,” Acta Polytechnica Hungarica, vol. 19, no. 2, pp. 43–57, 2022. and deep learning,” International Journal of Information Technology &
[54] I. K. Sastrawan, I. Bayupati, and D. M. S. Arsa, “Detection of fake news Decision Making, pp. 1–36, 2023.
using deep learning cnn–rnn based methods,” ICT Express, vol. 8, no. 3, [74] G. Mareeswari and E. V. Dinesh, “Deep neural networks based detection
pp. 396–408, 2022. and analysis of fake tweets,” in 2023 4th International Conference on
[55] K. L. Tan, C. P. Lee, and K. M. Lim, “Fn-net: A deep convolutional neural Signal Processing and Communication (ICSPC). IEEE, 2023, pp. 56–
network for fake news detection,” in 2021 9th International Conference 61.
on Information and Communication Technology (ICoICT). IEEE, 2021, [75] M. Samadi and S. Momtazi, “Fake news detection: deep semantic rep-
pp. 331–336. resentation with enhanced feature engineering,” International Journal of
[56] M. P. Thilakarathna, V. A. Wijayasekara, Y. Gamage, K. H. Peiris, Data Science and Analytics, pp. 1–12, 2023.
C. Abeysinghe, I. Rafaideen, and P. Vekneswaran, “Hybrid approach and [76] J. Alghamdi, Y. Lin, and S. Luo, “Does context matter? effective deep
architecture to detect fake news on twitter in real-time using neural net- learning approaches to curb fake news dissemination on social media,”
works,” in 2020 5th International Conference on Information Technology Applied Sciences, vol. 13, no. 5, p. 3345, 2023.
Research (ICITR). IEEE, 2020, pp. 1–6. [77] F. W. R. Tokpa, B. H. Kamagaté, V. Monsan, and S. Oumtanaga, “Fake
[57] F. Torgheh, M. R. Keyvanpour, B. Masoumi, and S. V. Shojaedini, “A news detection in social media: Hybrid deep learning approaches,” J.
novel method for detecting fake news: Deep learning based on propa- Adv. Inf. Technol, vol. 14, no. 3, pp. 606–615, 2023.
gation path concept,” in 2021 26th International Computer Conference, [78] O. Prakash and R. Kumar, “Fake news detection in social networks using
Computer Society of Iran (CSICC), IEEE. IEEE, 2021, pp. 1–5. attention mechanism,” in Proceedings of the International Conference on
VOLUME 4, 2016 21
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3435497
Cognitive and Intelligent Computing: ICCIC 2021, Volume 2. Springer, [99] S. Suratkar and F. Kazi, “Deep fake video detection using transfer
2023, pp. 453–462. learning approach,” Arabian Journal for Science and Engineering, vol. 48,
[79] S. Kumar, A. Kumar, A. Mallik, and R. R. Singh, “Optnet-fake: Fake no. 8, pp. 9727–9737, 2023.
news detection in socio-cyber platforms using grasshopper optimization [100] J. Del Ser, M. N. Bilbao, I. Laña, K. Muhammad, and D. Camacho,
and deep neural network,” IEEE Transactions on Computational Social “Efficient fake news detection using bagging ensembles of bidirectional
Systems, 2023. echo state networks,” in 2022 International Joint Conference on Neural
[80] Q. Zhang, Z. Guo, Y. Zhu, P. Vijayakumar, A. Castiglione, and B. B. Networks (IJCNN). IEEE, 2022, pp. 1–7.
Gupta, “A deep learning-based fast fake news detection model for cyber- [101] A. Sedik, A. A. Abohany, K. M. Sallam, K. Munasinghe, and T. Medhat,
physical social services,” Pattern Recognition Letters, vol. 168, pp. 31– “Deep fake news detection system based on concatenated and recurrent
38, 2023. modalities,” Expert Systems with Applications, vol. 208, p. 117953,
[81] A. Kishwar and A. Zafar, “Fake news detection on pakistani news using 2022.
machine learning and deep learning,” Expert Systems with Applications, [102] K. Sangeeta, T. G. Priya, S. Patro, A. Goutham, K. J. Jyothi, and M. A.
vol. 211, p. 118558, 2023. Kumar, “Fake news detection using feature selection and deep learning,”
[82] P. K. Verma, P. Agrawal, V. Madaan, and R. Prodan, “Mcred: multi-modal International Journal for Research in Applied Science & Engineering
message credibility for fake news detection using bert and cnn,” Journal Technology (IJRASET), vol. 10, pp. 3878–3887, 2022.
of Ambient Intelligence and Humanized Computing, vol. 14, no. 8, pp. [103] J. Rautela, V. Ramalingam, and H. Makhdoomi, “Fake news detection
10 617–10 629, 2023. through deep learning techniques,” International Journal of Health Sci-
[83] A. K. Yadav, S. Kumar, D. Kumar, L. Kumar, K. Kumar, S. K. Maurya, ences, vol. 6, no. 5, pp. 2107––2111, 2022.
M. Kumar, and D. Yadav, “Fake news detection using hybrid deep [104] V. Kandasamy, Š. Hubálovskỳ, and P. Trojovskỳ, “Deep fake detection
learning method,” SN Computer Science, vol. 4, no. 6, p. 845, 2023. using a sparse auto encoder with a graph capsule dual graph cnn,” PeerJ
[84] A. Saeed and E. Al Solami, “Fake news detection using machine learn- Computer Science, vol. 8, p. e953, 2022.
ing and deep learning methods,” CMC-COMPUTERS MATERIALS & [105] Y. Tashtoush, B. Alrababah, O. Darwish, M. Maabreh, and N. Alsaedi, “A
CONTINUA, vol. 77, no. 2, pp. 2079–2096, 2023. deep learning framework for detection of covid-19 fake news on social
[85] A. Y. Umar, I. S. Ahmad, and K. Muhammad, “Fake news detection using media platforms,” Data, vol. 7, no. 5, p. 65, 2022.
cnn/gru deep learning model,” Sule Lamido University Journal of Science [106] R. Muppidi and V. Biksham, “Deep convolutional neural network for
& Technology, vol. 7, no. 1, pp. 57–67, 2023. fake news detection over online social networks,” INTERANTIONAL
JOURNAL OF SCIENTIFIC RESEARCH IN ENGINEERING AND
[86] Y. Singh and P. Singh, “Fake news detection using lstm in tensorflow and
MANAGEMENT, vol. 6, no. 4, pp. 1–10, 2022.
deep learning,” Journal of Applied Science and Education (JASE), vol. 3,
no. 2, pp. 1–14, 2023. [107] Q. Li, Q. Hu, Y. Lu, Y. Yang, and J. Cheng, “Multi-level word features
based on cnn for fake news detection in cultural communication,” Per-
[87] P. Sharma and R. Sahu, “Fake news detection using deep learning
sonal and Ubiquitous Computing, vol. 24, pp. 259–272, 2020.
based approach,” in 2023 International Conference on Circuit Power and
[108] S. Deepak and B. Chitturi, “Deep neural approach to fake-news identifi-
Computing Technologies (ICCPCT). IEEE, 2023, pp. 651–656.
cation,” Procedia Computer Science, vol. 167, pp. 2236–2243, 2020.
[88] M. R. H. Shezan, M. N. Zawad, Y. A. Shahed, and S. Ripon, “Bangla
[109] C. Kulkarni, P. Monika, S. Shruthi, M. Deepak Bharadwaj, and D. Uday,
fake news detection using hybrid deep learning models,” in Applied
“Covid-19 fake news detection using glove and bi-lstm,” in Proceedings
Informatics for Industry 4.0. Chapman and Hall/CRC, 2023, pp. 46–
of Second International Conference on Sustainable Expert Systems:
60.
ICSES 2021. Springer, 2022, pp. 43–56.
[89] C. Nandhakumar, C. Kowsika, R. Reshema, and L. Sandhiya, “Fake news
[110] R. Malhotra, A. Mahur et al., “Covid-19 fake news detection system,” in
detection using machine learning and deep learning classifiers,” in Inter-
2022 12th International Conference on Cloud Computing, Data Science
national Conference on Information and Communication Technology for
& Engineering (Confluence). IEEE, 2022, pp. 428–433.
Intelligent Systems. Springer, 2023, pp. 165–175.
[111] Y. Tian, J. Gu, Y. Jia, and R. O. Sinnott, “An exploration of machine
[90] P. M. Subhash, D. Gupta, S. Palaniswamy, and M. Venugopalan, “Fake and deep learning models for fake news detection in social media,” in
news detection using deep learning and transformer-based model,” in 2021 8th International Conference on Behavioral and Social Computing
2023 14th International Conference on Computing Communication and (BESC). IEEE, 2021, pp. 1–6.
Networking Technologies (ICCCNT). IEEE, 2023, pp. 1–6.
[112] H. Zhu and R. O. Sinnott, “A performance comparison of fake news de-
[91] A. Jaiswal, H. Verma, and N. Sachdeva, “Swarm optimized fake news tection approaches,” in 2021 IEEE Asia-Pacific Conference on Computer
detection on social-media textual content using deep learning,” in 2023 Science and Data Engineering (CSDE). IEEE, 2021, pp. 1–7.
International Conference on Advances in Computing, Communication [113] G. S. Mahara and S. Gangele, “Fake news detection: A rnn-lstm, bi-
and Applied Informatics (ACCAI). IEEE, 2023, pp. 1–8. lstm based deep learning approach,” in 2022 IEEE 1st International
[92] I. Ennejjai, A. Ariss, N. Kharmoum, W. Rhalem, S. Ziti, and Conference on Data, Decision and Systems (ICDDS). IEEE, 2022, pp.
M. Ezziyyani, “Artificial intelligence for fake news,” in International 01–06.
Conference on Advanced Intelligent Systems for Sustainable Develop- [114] G. Anusha, G. Praveen, D. Mounika, U. S. Krishna, and R. Cristin,
ment. Springer, 2022, pp. 77–91. “Detection of fake news using recurrent neural network,” in 2022 IEEE
[93] O. Ngada and B. Haskins, “Investigating fake news detection by means International Conference on Distributed Computing and Electrical Cir-
of deep learning on a limited data set,” in 2022 IEEE Asia-Pacific cuits and Electronics (ICDCECE). IEEE, 2022, pp. 1–5.
Conference on Computer Science and Data Engineering (CSDE). IEEE, [115] P. K. Sree, G. R. Babu, P. R. Rao, P. V. Chintalapati, M. Prasad et al.,
2022, pp. 1–6. “Fake news detection using cellular automata based deep learning,”
[94] Z. Wang, “Deep learning methods for fake news detection,” in 2022 IEEE in 2023 3rd International Conference on Computing and Information
2nd International Conference on Data Science and Computer Application Technology (ICCIT). IEEE, 2023, pp. 167–171.
(ICDSCA). IEEE, 2022, pp. 472–475. [116] H. Ali, M. Khan, A. AlGhadhban, M. Alazmi, A. Alzamil, K. Al-utaibi,
[95] N. Jayakody, A. Mohammad, and M. N. Halgamuge, “Fake news detec- and J. Qadir, “Analyzing the robustness of fake-news detectors under
tion using a decentralized deep learning model and federated learning,” in black-box adversarial attacks,” IEEE Access, vol. 9, pp. 81 678–81 692,
IECON 2022–48th Annual Conference of the IEEE Industrial Electronics 2021.
Society. IEEE, 2022, pp. 1–6. [117] E. Qawasmeh, M. Tawalbeh, and M. Abdullah, “Automatic identification
[96] D. R. Collen, L. K. Nyandoro, and K. Zvarevashe, “Fake news detection of fake news using deep learning,” in 2019 Sixth international confer-
using 5l-cnn,” in 2022 1st Zimbabwe Conference of Information and ence on social networks analysis, Management and Security (SNAMS).
Communication Technologies (ZCICT). IEEE, 2022, pp. 1–7. IEEE, 2019, pp. 383–388.
[97] A. Qdroo and M. Baykara, “A new approach to detect fake news related [118] K. K. Kumar, S. H. Rao, G. Srikar, and M. B. Chandra, “A novel
to covid-19 pandemic using deep neural network,” Journal of Applied approach for detection of fake news using long short term memory
Science and Technology Trends, vol. 3, no. 02, pp. 27–34, 2022. (lstm),” International Journal, vol. 10, no. 5, 2021.
[98] Z. A. Jawad and A. J. Obaid, “Combination of convolution neural net- [119] T. Ahmad, M. S. Faisal, A. Rizwan, R. Alkanhel, P. W. Khan, and
works and deep neural networks for fake news detection,” arXiv preprint A. Muthanna, “Efficient fake news detection mechanism using enhanced
arXiv:2210.08331, 2022. deep learning model,” Applied Sciences, vol. 12, no. 3, p. 1743, 2022.
22 VOLUME 4, 2016
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3435497
[120] N. Aslam, I. Ullah Khan, F. S. Alotaibi, L. A. Aldaej, and A. K. national Conference on Knowledge, Information and Creativity Support
Aldubaikil, “Fake detect: A deep learning ensemble model for fake news Systems (KICSS). IEEE, 2018, pp. 1–5.
detection,” complexity, vol. 2021, pp. 1–8, 2021. [141] P. Shrivastava and D. K. Sharma, “Fake content identification using
[121] T. Chauhan and H. Palivela, “Optimization and improvement of fake pre-trained glove-embedding,” in 2021 5th International Conference on
news detection using deep learning approaches for societal benefit,” Information Systems and Computer Networks (ISCON). IEEE, 2021,
International Journal of Information Management Data Insights, vol. 1, pp. 1–6.
no. 2, p. 100051, 2021. [142] T. E. Trueman, A. Kumar, P. Narayanasamy, and J. Vidya, “Attention-
[122] M.-Y. Chen, Y.-W. Lai, and J.-W. Lian, “Using deep learning models based c-bilstm for fake news detection,” Applied Soft Computing, vol.
to detect fake news about covid-19,” ACM Transactions on Internet 110, p. 107600, 2021.
Technology, 2022. [143] M. Umer, Z. Imtiaz, S. Ullah, A. Mehmood, G. S. Choi, and B.-W. On,
[123] R. Garg and S. Jeevaraj, “Effective fake news classifier and its applica- “Fake news stance detection using deep learning architecture (cnn-lstm),”
tions to covid-19,” in 2021 IEEE Bombay Section Signature Conference IEEE Access, vol. 8, pp. 156 695–156 706, 2020.
(IBSSC). IEEE, 2021, pp. 1–6. [144] P. Ushashree, A. Naik, S. Gurav, A. Kumar, S. Chethan, and B. Madhu-
[124] D. K. Jain, A. Kumar, and A. Shrivastava, “Canardeep: a hybrid deep mala, “Fake news detection using neural network,” in 2023 IEEE Inter-
neural model with mixed fusion for rumour detection in social data national Conference on Integrated Circuits and Communication Systems
streams,” Neural Computing and Applications, pp. 1–12, 2022. (ICICACS). IEEE, 2023, pp. 01–05.
[125] V. Jain, R. K. Kaliyar, A. Goswami, P. Narang, and Y. Sharma, “Aenet: [145] A. Anand, R. Kulkarni, and P. Agrawal, “Fake news identification:
an attention-enabled neural architecture for fake news detection using An effective combined approach using ml and dl techniques,” in 2023
contextual features,” Neural Computing and Applications, vol. 34, no. 1, 2nd International Conference on Paradigm Shifts in Communications
pp. 771–782, 2022. Embedded Systems, Machine Learning and Signal Processing (PCEMS).
[126] T. Jiang, J. P. Li, A. U. Haq, and A. Saboor, “Fake news detection using IEEE, 2023, pp. 1–6.
deep recurrent neural networks,” in 2020 17th International Computer [146] K. Saini and R. Jain, “A hybrid lstm-bert and glove-based deep learning
Conference on Wavelet Active Media Technology and Information Pro- approach for the detection of fake news,” in 2023 3rd International
cessing (ICCWAMTIP). IEEE, 2020, pp. 205–208. Conference on Smart Data Intelligence (ICSMDI). IEEE, 2023, pp.
[127] N. Kanagavalli, S. B. Priya, and D. Jeyakumar, “Design of hyperparam- 400–406.
eter tuned deep learning based automated fake news detection in social [147] A. Khoudi, N. Yahiaoui, and F. Rebahi, “Detect misinformation of covid-
networking data,” in 2022 6th International Conference on Computing 19 using deep learning: A comparative study based on word embedding,”
Methodologies and Communication (ICCMC). IEEE, 2022, pp. 958– in 2023 1st International Conference on Advanced Innovations in Smart
963. Cities (ICAISC). IEEE, 2023, pp. 1–5.
[128] J. Kumari, R. Choudhary, S. Kumari, and G. Krishna, “A deep learning [148] S. M. Bankar and S. K. Gupta, “Fake news detection using lstm-based
based approach for classification of news as real or fake,” in Data Science deep learning approach and word embedding feature extraction,” in
and Security: Proceedings of IDSCS 2021. Springer, 2021, pp. 239–246. International Conference on Communication, Electronics and Digital
Technology. Springer, 2023, pp. 129–141.
[129] D.-H. Lee, Y.-R. Kim, H.-J. Kim, S.-M. Park, and Y.-J. Yang, “Fake
[149] A. Matheven and B. V. D. Kumar, “Fake news detection using deep
news detection using deep learning,” Journal of Information Processing
learning and natural language processing,” in 2022 9th International
Systems, vol. 15, no. 5, pp. 1119–1130, 2019.
Conference on Soft Computing & Machine Intelligence (ISCMI). IEEE,
[130] J. Liu, C. Wang, C. Li, N. Li, J. Deng, and J. Z. Pan, “Dtn: Deep
2022, pp. 11–14.
triple network for topic specific fake news detection,” Journal of Web
[150] J. A. Reshi and R. Ali, “Online fake news detection using pre-trained em-
Semantics, vol. 70, p. 100646, 2021.
beddings,” in 2022 5th International Conference on Multimedia, Signal
[131] R. Mahesh, B. Poornika, N. Sharaschandrika, S. D. Goud, and P. U.
Processing and Communication Technologies (IMPACT). IEEE, 2022,
Kumar, “Identification of fake news using deep learning architecture,”
pp. 1–5.
in 2021 Third International Conference on Inventive Research in Com-
[151] M. Madani, H. Motameni, and H. Mohamadi, “Fake news detection using
puting Applications (ICIRCA). IEEE, 2021, pp. 1246–1253.
deep learning integrating feature extraction, natural language processing,
[132] B. Majumdar, M. RafiuzzamanBhuiyan, M. A. Hasan, M. S. Islam, and and statistical descriptors,” Security and Privacy, vol. 5, no. 6, p. e264,
S. R. H. Noori, “Multi class fake news detection using lstm approach,” 2022.
in 2021 10th International Conference on System Modeling & Advance- [152] P. Katariya, V. Gupta, R. Arora, A. Kumar, S. Dhingra, Q. Xin, and
ment in Research Trends (SMART). IEEE, 2021, pp. 75–79. J. Hemanth, “A deep neural network-based approach for fake news
[133] S. Mengji, S. Ambarte, S. V. T. Arumilli, S. Mhamane, and R. Rane, detection in regional language,” International Journal of Web Information
“Fake news detection using rnn-lstm,” International Journal for Research Systems, vol. 18, no. 5/6, pp. 286–309, 2022.
in Applied Science & Engineering Technology (IJRASET), vol. 9, pp. [153] A. Divija, S. K. Kurkuri, P. Telukuntla, and S. Mantha, “Fake news
1731–1737, 2021. classifier,” International Journal for Research in Applied Science &
[134] A. Mohapatra, N. Thota, and P. Prakasam, “Fake news detection and Engineering Technology (IJRASET), vol. 10, pp. 1716–1722, 2022.
classification using hybrid bilstm and self-attention model,” Multimedia [154] J. Soni, “An efficient lstm model for fake news detection,” Computer
Tools and Applications, vol. 81, no. 13, pp. 18 503–18 519, 2022. Science & Engineering: An International Journal (CSEIJ), vol. 12, no. 2,
[135] U. Narayan, A. Kumar, and K. Kumar, “Fake news detection using hybrid 2022.
of deep neural network and stacked lstm,” in 2021 3rd International [155] E. Amer, K.-S. Kwak, and S. El-Sappagh, “Context-based fake news
Conference on Advances in Computing, Communication Control and detection model relying on deep learning models,” Electronics, vol. 11,
Networking (ICAC3N). IEEE, 2021, pp. 385–390. no. 8, p. 1255, 2022.
[136] N. Rai, D. Kumar, N. Kaushik, C. Raj, and A. Ali, “Fake news clas- [156] N. Xiang et al., “Deep learning-based fake information detection and
sification using transformer based enhanced lstm and bert,” International influence evaluation,” Computational Intelligence and Neuroscience, vol.
Journal of Cognitive Computing in Engineering, vol. 3, pp. 98–105, 2022. 2022, 2022.
[137] R. Rajalaxmi, L. Narasimha Prasad, B. Janakiramaiah, C. Pavankumar, [157] S. M. Jaybhaye, V. Badade, A. Dodke, A. Holkar, and P. Lokhande, “Fake
N. Neelima, and V. Sathishkumar, “Optimizing hyperparameters and news detection using lstm based deep learning approach,” in ITM Web of
performance analysis of lstm model in detecting fake news on social Conferences, vol. 56. EDP Sciences, 2023, p. 03005.
media,” Transactions on Asian and Low-Resource Language Information [158] R. S. Aziz, A. T. Sadiq, M. Kherallah, and A. Douik, “Arabic fake
Processing, 2022. news detection for covid-19 using deep learning and machine learning,”
[138] F. Sadeghi, A. J. Bidgoly, and H. Amirkhani, “Fake news detection on Periodicals of Engineering and Natural Sciences, vol. 11, no. 6, pp. 56–
social media using a natural language inference approach,” Multimedia 72, 2023.
Tools and Applications, vol. 81, no. 23, pp. 33 801–33 821, 2022. [159] M. B. Narayanan, A. K. Ramesh, K. Gayathri, and A. Shahina, “Fake
[139] S. R. Sahoo and B. B. Gupta, “Multiple features based approach for news detection using a deep learning transformer based encoder-decoder
automatic fake news detection on social networks using deep learning,” architecture,” Journal of Intelligent & Fuzzy Systems, no. Preprint, pp.
Applied Soft Computing, vol. 100, p. 106983, 2021. 1–13, 2023.
[140] Y. Seo and C.-S. Jeong, “Fagon: Fake news detection model using [160] S. Malik, A. K. Chakraverti, and A. I. Abidi, “Enhancing fake news
grammatic transformation on neural network,” in 2018 Thirteenth Inter- detection using classification algorithms and deep learning,” in 2023
VOLUME 4, 2016 23
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3435497
10th IEEE Uttar Pradesh Section International Conference on Electrical, PROPOR 2020, Evora, Portugal, March 2–4, 2020, Proceedings 14.
Electronics and Computer Engineering (UPCON), vol. 10. IEEE, 2023, Springer, 2020, pp. 143–152.
pp. 780–787. [179] A. Al Obaid, H. Khotanlou, M. Mansoorizadeh, and D. Zabihzadeh,
[161] A. Chabukswar, P. D. Shenoy, and K. Venugopal, “Fake news detection “Multimodal fake-news recognition using ensemble of deep learners,”
using optimized deep learning model through effective feature extrac- Entropy, vol. 24, no. 9, p. 1242, 2022.
tion,” in 2023 International Conference on Recent Advances in Informa- [180] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training
tion Technology for Sustainable Development (ICRAIS). IEEE, 2023, of deep bidirectional transformers for language understanding,” arXiv
pp. 118–123. preprint arXiv:1810.04805, 2018.
[162] N. Ahuja and S. Kumar, “S-han: Hierarchical attention networks with [181] S. M. Isa, G. Nico, and M. Permana, “Indobert for indonesian fake news
stacked gated recurrent unit for fake news detection,” in 2020 8th Interna- detection,” ICIC Express Letters, vol. 16, no. 3, pp. 289–297, 2022.
tional Conference on Reliability, Infocom Technologies and Optimization [182] B. Palani, S. Elango, and V. Viswanathan K, “Cb-fake: A multimodal
(Trends and Future Directions)(ICRITO). IEEE, 2020, pp. 873–877. deep learning framework for automatic fake news detection using capsule
[163] G. Sarin and P. Kumar, “Convgrutext: a deep learning method for fake neural network and bert,” Multimedia Tools and Applications, vol. 81,
text detection on online social media,” Pasific Asia Conference on Infor- no. 4, pp. 5587–5620, 2022.
mation Systems, PACIS 2020 Proceedings, 2020. [183] S. Sharma, M. Saraswat, and A. K. Dubey, “Fake news detection us-
[164] A. Verma, V. Mittal, and S. Dawn, “Find: Fake information and news ing deep learning,” in Knowledge Graphs and Semantic Web: Third
detections using deep learning,” in 2019 twelfth international conference Iberoamerican Conference and Second Indo-American Conference,
on contemporary computing (IC3). IEEE, 2019, pp. 1–7. KGSWC 2021, Kingsville, Texas, USA, November 22–24, 2021, Pro-
[165] M. Dong, L. Yao, X. Wang, B. Benatallah, Q. Z. Sheng, and H. Huang, ceedings 3. Springer, 2021, pp. 249–259.
“Dual: A deep unified attention model with latent relation representations [184] M. Kanchana, V. M. Kumar, T. Anish, and P. Gopirajan, “Deep fake
for fake news detection,” in Web Information Systems Engineering– bert: Efficient online fake news detection system,” in 2023 International
WISE 2018: 19th International Conference, Dubai, United Arab Emi- Conference on Networking and Communications (ICNWC). IEEE,
rates, November 12-15, 2018, Proceedings, Part I 19. Springer, 2018, 2023, pp. 1–6.
pp. 199–209. [185] R. H. Khan, A. Shihavuddin, M. M. Syeed, R. U. Haque, and M. F. Uddin,
[166] S. Taheri, S. H. Hashemi, A. Y. Zomaya, and J. Yong, “Sequence graph “Improved fake news detection method based on deep learning and
transform: A general approach to substructure-aware sequence encod- comparative analysis with other machine learning approaches,” in 2022
ing,” in 2018 IEEE International Conference on Data Mining (ICDM). International Conference on Engineering and Emerging Technologies
IEEE, 2018, pp. 117–126. (ICEET). IEEE, 2022, pp. 1–6.
[167] P. V. G. C. A. Casanova, A. R. P. Lio, and Y. Bengio, “Graph attention [186] A. Kumar, J. P. Singh, and A. K. Singh, “Covid-19 fake news detection
networks,” ICLR. Petar Velickovic Guillem Cucurull Arantxa Casanova using ensemble-based deep learning model,” IT Professional, vol. 24,
Adriana Romero Pietro Liò and Yoshua Bengio, 2018. no. 2, pp. 32–37, 2022.
[168] W. Hamilton, Z. Ying, and J. Leskovec, “Inductive representation learn- [187] R. K. Kaliyar, A. Goswami, and P. Narang, “Fakebert: Fake news
ing on large graphs,” Advances in neural information processing systems, detection in social media with a bert-based deep learning approach,”
vol. 30, 2017. Multimedia tools and applications, vol. 80, no. 8, pp. 11 765–11 788,
[169] T. N. Kipf and M. Welling, “Semi-supervised classification with graph 2021.
convolutional networks,” arXiv preprint arXiv:1609.02907, 2016. [188] M. Q. Alnabhan and P. Branco, “Evaluating deep learning for cross-
[170] F. Scarselli, M. Gori, A. C. Tsoi, M. Hagenbuchner, and G. Monfar- domains fake news detection,” in International Symposium on Founda-
dini, “The graph neural network model,” IEEE transactions on neural tions and Practice of Security. Springer, 2023, pp. 40–51.
networks, vol. 20, no. 1, pp. 61–80, 2008. [189] J. V. Tembhurne, M. M. Almin, and T. Diwan, “Mc-dnn: Fake news de-
[171] N. Bai, F. Meng, X. Rui, and Z. Wang, “Rumour detection based on graph tection using multi-channel deep neural networks,” International Journal
convolutional neural net,” IEEE Access, vol. 9, pp. 21 686–21 693, 2021. on Semantic Web and Information Systems (IJSWIS), vol. 18, no. 1, pp.
[172] F. B. Mahmud, M. M. S. Rayhan, M. H. Shuvo, I. Sadia, and M. K. 1–20, 2022.
Morol, “A comparative analysis of graph neural networks and commonly [190] R. K. Kaliyar, P. Kumar, M. Kumar, M. Narkhede, S. Namboodiri, and
used machine learning algorithms on fake news detection,” in 2022 S. Mishra, “Deepnet: an efficient neural network for fake news detection
7th International Conference on Data Science and Machine Learning using news-user engagements,” in 2020 5th International Conference on
Applications (CDMA). IEEE, 2022, pp. 97–102. Computing, Communication and Security (ICCCS). IEEE, 2020, pp.
[173] I. Pilkevych, D. Fedorchuk, O. Naumchak, and M. Romanchuk, “Fake 1–6.
news detection in the framework of decision-making system through [191] F. Zhou, Y. Hu, and X. Shen, “Msanet: multimodal self-augmentation and
graph neural network,” in 2021 IEEE 4th International Conference on Ad- adversarial network for rgb-d object recognition,” The Visual Computer,
vanced Information and Communication Technologies (AICT). IEEE, vol. 35, no. 11, pp. 1583–1594, 2019.
2021, pp. 153–157. [192] A. Mumuni and F. Mumuni, “Data augmentation: A comprehensive
[174] Y. Ren, B. Wang, J. Zhang, and Y. Chang, “Adversarial active learning survey of modern approaches,” Array, vol. 16, p. 100258, 2022.
based heterogeneous graph neural network for fake news detection,” in [193] S. K. Hamed, M. J. Ab Aziz, and M. R. Yaakub, “A review of fake news
2020 IEEE International Conference on Data Mining (ICDM). IEEE, detection models: Highlighting the factors affecting model performance
2020, pp. 452–461. and the prominent techniques used,” International Journal of Advanced
[175] M. Sun, I. A. Hameed, H. Wang, and M. Pasquine, “Perceiving the Computer Science and Applications, vol. 14, no. 7, 2023.
narrative style for fake news detection using deep learning,” in 2021 IEEE [194] I. Ahmad, M. Yousaf, S. Yousaf, and M. O. Ahmad, “Fake news detection
23rd Int Conf on High Performance Computing & Communications; 7th using machine learning ensemble methods,” Complexity, vol. 2020, pp.
Int Conf on Data Science & Systems; 19th Int Conf on Smart City; 1–11, 2020.
7th Int Conf on Dependability in Sensor, Cloud & Big Data Systems [195] A. Zubiaga, M. Liakata, and R. Procter, “Learning reporting dynamics
& Application (HPCC/DSS/SmartCity/DependSys). IEEE, 2021, pp. during breaking news for rumour detection in social media,” arXiv
1195–1202. preprint arXiv:1610.07363, 2016.
[176] B. Upadhayay and V. Behzadan, “Hybrid deep learning model for fake [196] W. Y. Wang, “" liar, liar pants on fire": A new benchmark dataset for fake
news detection in social networks (student abstract),” in Proceedings of news detection,” arXiv preprint arXiv:1705.00648, 2017.
the AAAI Conference on Artificial Intelligence, vol. 36, no. 11, 2022, pp. [197] K. Shu, D. Mahudeswaran, S. Wang, D. Lee, and H. Liu, “Fakenewsnet:
13 067–13 068. A data repository with news content, social context, and spatiotemporal
[177] P. Hiremath, S. S. Kalagi et al., “Analysis of fake news detection using information for studying fake news on social media,” Big data, vol. 8,
graph neural network (gnn) and deep learning,” in 2023 2nd Interna- no. 3, pp. 171–188, 2020.
tional Conference on Automation, Computing and Renewable Systems [198] M. Nirav Shah and A. Ganatra, “A systematic literature review and
(ICACRS). IEEE, 2023, pp. 1805–1811. existing challenges toward fake news detection models,” Social Network
[178] E. Y. Okano, Z. Liu, D. Ji, and E. E. S. Ruiz, “Fake news detection Analysis and Mining, vol. 12, no. 1, p. 168, 2022.
on fake. br using hierarchical attention networks,” in Computational [199] B. Cao, L. Hua, J. Cao, J. Gui, B. Liu, and J. T.-Y. Kwok, “No place
Processing of the Portuguese Language: 14th International Conference, to hide: Dual deep interaction channel network for fake news detection
24 VOLUME 4, 2016
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3435497
based on data augmentation,” arXiv preprint arXiv:2303.18049, vol. 14, [222] Microsoft, “Microsoft machine learning,” Year of the version or release.
no. 8, pp. 1–10, 2023. [Online]. Available: https://docs.microsoft.com/en-us/machine-learning/
[200] S. Warjri, P. Pakray, S. A. Lyngdoh, and A. K. Maji, “Fake news detection [223] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification
using social media data for khasi language,” in 2023 International Confer- with deep convolutional neural networks,” Communications of the ACM,
ence on Intelligent Systems, Advanced Computing and Communication vol. 60, no. 6, pp. 84–90, 2012.
(ISACC). IEEE, 2023, pp. 1–6. [224] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
[201] S. Helmstetter and H. Paulheim, “Collecting a large scale dataset for large-scale image recognition,” arXiv preprint arXiv:1409.1556, 2014.
classifying fake news tweets using weak supervision,” Future Internet, [225] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image
vol. 13, no. 5, p. 114, 2021. recognition,” Proceedings of the IEEE conference on computer vision and
[202] K. D. K. Parimala and A. G. Mala, “An optimal detection of fake news pattern recognition (CVPR), pp. 770–778, 2016.
from twitter data using dual-stage deep capsule autoencoder,” Journal [226] T. Mikolov, K. Chen, G. Corrado, and J. Dean, “Efficient estimation of
of Experimental & Theoretical Artificial Intelligence, vol. 36, no. 2, pp. word representations in vector space,” arXiv preprint arXiv:1301.3781,
287–313, 2024. 2013.
[203] S. Kato, L. Yang, and D. Ikeda, “Domain bias in fake news datasets [227] J. Pennington, R. Socher, and C. D. Manning, “Glove: Global vectors for
consisting of fake and real news pairs,” in 2022 12th International word representation,” Empirical Methods in Natural Language Process-
Congress on Advanced Applied Informatics (IIAI-AAI), 2022, pp. 101– ing (EMNLP), pp. 1532–1543, 2014.
106. [228] W. Ferreira and A. Vlachos, “Emergent: a novel data-set for stance
[204] H. F. Villela, F. Corrêa, J. S. d. A. N. Ribeiro, A. Rabelo, and D. B. F. classification,” in Proceedings of NAACL-HLT, 2016, pp. 1163–1168.
Carvalho, “Fake news detection: a systematic literature review of ma- [229] B. Palani and S. Elango, “Ctrl-fnd: content-based transfer learning ap-
chine learning algorithms and datasets,” Journal on Interactive Systems, proach for fake news detection on social media,” International Journal
vol. 14, no. 1, pp. 47–58, 2023. of System Assurance Engineering and Management, vol. 14, no. 3, pp.
[205] S. Suryavardan, S. Mishra, M. Chakraborty, P. Patwa, A. Rani, 903–918, 2023.
A. Chadha, A. Reganti, A. Das, A. Sheth, M. Chinnakotla et al., [230] B. M. Pavlyshenko, “Analysis of disinformation and fake news
“Findings of factify 2: multimodal fake news detection,” arXiv preprint detection using fine-tuned large language model,” arXiv preprint
arXiv:2307.10475, 2023. arXiv:2309.04704, 2023.
[206] K. Weiss, T. M. Khoshgoftaar, and D. Wang, “A survey of transfer [231] C. Seiffert, T. M. Khoshgoftaar, J. Van Hulse, and A. Napolitano, “A
learning,” Journal of Big data, vol. 3, no. 1, pp. 1–40, 2016. comparative study of data sampling and cost sensitive learning,” in 2008
[207] S. J. Pan and Q. Yang, “A survey on transfer learning,” IEEE Transactions IEEE international conference on data mining workshops. IEEE, 2008,
on knowledge and data engineering, vol. 22, no. 10, pp. 1345–1359, 2010. pp. 46–52.
[208] C. Wang and S. Mahadevan, “Heterogeneous domain adaptation using [232] R. Longadge and S. Dongre, “Class imbalance problem in data mining
manifold alignment,” in IJCAI proceedings-international joint conference review,” arXiv preprint arXiv:1305.1707, 2013.
on artificial intelligence, vol. 22, no. 1, 2011, p. 1541. [233] P. Branco, L. Torgo, and R. P. Ribeiro, “A survey of predictive modeling
[209] Y. Zhu, Y. Chen, Z. Lu, S. Pan, G.-R. Xue, Y. Yu, and Q. Yang, “Het- on imbalanced domains,” ACM computing surveys (CSUR), vol. 49,
erogeneous transfer learning for image classification,” in Proceedings of no. 2, pp. 1–50, 2016.
the AAAI conference on artificial intelligence, vol. 25, no. 1, 2011, pp. [234] N. V. Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer,
1304–1309. “Smote: synthetic minority over-sampling technique,” Journal of artificial
[210] B. Kulis, K. Saenko, and T. Darrell, “What you saw is not what you get: intelligence research, vol. 16, pp. 321–357, 2002.
Domain adaptation using asymmetric kernel transforms,” in CVPR 2011. [235] H. He, Y. Bai, E. A. Garcia, and S. Li, “Adasyn: Adaptive synthetic sam-
IEEE, 2011, pp. 1785–1792. pling approach for imbalanced learning,” in 2008 IEEE international joint
[211] F. Arslan, N. Hassan, C. Li, and M. Tremayne, “A benchmark dataset of conference on neural networks (IEEE world congress on computational
check-worthy factual claims,” in Proceedings of the International AAAI intelligence). IEEE, 2008, pp. 1322–1328.
Conference on Web and Social Media, vol. 14, 2020, pp. 821–829. [236] N. V. Chawla, N. Japkowicz, and A. Kotcz, “Special issue on learning
[212] M. Harel and S. Mannor, “Learning from multiple outlooks,” arXiv from imbalanced data sets,” ACM SIGKDD explorations newsletter,
preprint arXiv:1005.0027, 2010. vol. 6, no. 1, pp. 1–6, 2004.
[213] J. Nam and S. Kim, “Heterogeneous defect prediction,” in Proceedings [237] T. Bhatia, B. Manaskasemsak, and A. Rungsawang, “Detecting fake news
of the 2015 10th joint meeting on foundations of software engineering, sources on twitter using deep neural network,” in 2023 11th international
2015, pp. 508–519. conference on information and education technology (ICIET). IEEE,
[214] P. Prettenhofer and B. Stein, “Cross-language text classification using 2023, pp. 508–512.
structural correspondence learning,” in Proceedings of the 48th annual [238] K. Ghosh, C. Bellinger, R. Corizzo, P. Branco, B. Krawczyk, and
meeting of the association for computational linguistics, 2010, pp. 1118– N. Japkowicz, “The class imbalance problem in deep learning,” Machine
1127. Learning, pp. 1–57, 2022.
[215] J. Yosinski, J. Clune, Y. Bengio, and H. Lipson, “How transferable [239] J.-G. Gaudreault, P. Branco, and J. Gama, “An analysis of performance
are features in deep neural networks?” Advances in neural information metrics for imbalanced classification,” in Discovery Science: 24th In-
processing systems, vol. 27, 2014. ternational Conference, DS 2021, Halifax, NS, Canada, October 11–13,
[216] Z. Alyafeai, M. S. AlShaibani, and I. Ahmad, “A survey on transfer learn- 2021, Proceedings 24. Springer, 2021, pp. 67–77.
ing in natural language processing,” arXiv preprint arXiv:2007.04239, [240] Z. Yan, Y. Zhang, X. Yuan, S. Lyu, and B. Wu, “Deepfakebench: A
2020. comprehensive benchmark of deepfake detection,” Advances in Neural
[217] J. Howard and S. Ruder, “Universal language model fine-tuning for text Information Pro-cessing Systems, vol. 2, no. 6, pp. 1–32, 2023.
classification,” arXiv preprint arXiv:1801.06146, 2018. [241] P. Shrivastava and D. K. Sharma, “Covid-19 fake news detection using
[218] M. Masum and H. Shahriar, “A transfer learning with deep neural pre-tuned bert-based transfer learning models,” in 2022 11th International
network approach for network intrusion detection,” International journal Conference on System Modeling & Advancement in Research Trends
of intellligent computing research, vol. 12, no. 1, 2021. (SMART). IEEE, 2022, pp. 64–68.
[219] C. Käding, E. Rodner, A. Freytag, and J. Denzler, “Fine-tuning deep [242] W. Tang, Z. Ma, H. Sun, and J. Wang, “Learning sparse alignments
neural networks in continuous learning scenarios,” in Computer Vision– via optimal transport for cross-domain fake news detection,” in ICASSP
ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, 2023-2023 IEEE International Conference on Acoustics, Speech and
Taiwan, November 20-24, 2016, Revised Selected Papers, Part III 13. Signal Processing (ICASSP). IEEE, 2023, pp. 1–5.
Springer, 2017, pp. 588–605. [243] I. Y. Agarwal and D. P. Rana, “Fake news and imbalanced data per-
[220] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, spective,” in Data Preprocessing, Active Learning, and Cost Perceptive
D. Erhan, V. Vanhoucke, and A. Rabinovich, “Going deeper with Approaches for Resolving Data Imbalance. IGI Global, 2021, pp. 195–
convolutions,” Google Research, 2015. [Online]. Available: https: 210.
//arxiv.org/abs/1409.4842 [244] A. J. Keya, M. A. H. Wadud, M. Mridha, M. Alatiyyah, and M. A.
[221] Microsoft Corporation, “Microsoftml: A package for machine learning Hamid, “Augfake-bert: handling imbalance through augmentation of fake
with r,” Version 1.5.0, 2018. [Online]. Available: https://microsoft. news using bert to enhance the performance of fake news classification,”
github.io/MicrosoftML/ Applied Sciences, vol. 12, no. 17, p. 8398, 2022.
VOLUME 4, 2016 25
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3435497
26 VOLUME 4, 2016
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4