Multimodal fake news detection model using improved heuristic approach with optimal weighted integration and dilated adaptive deep learning

Abstract

A lot of fake news is published with the help of television and social media. As a consequence, we are affected by the disinformation and misinformation spreading in our community. It is implemented to identify fake news creators to prevent spreading misinformation about the people. Thus, it is published like real news for the reputation and finances of an individual. It is challenging, because that is created by the integration of real and false details, and images are attached like originals to confuse the public. Few fake news detection tools are available to detect fake information. To solve this fake news spreading problem, an effectual multimodal fake news detection system is proposed based on the deep learning technique. In the beginning, the input image and text are gathered from benchmark data sources. Consequently, the deep features are extracted from raw images and text. The dilated Visual Geometry Group 16 (VGG16) is adopted to extract the image features and similarly, the text features are retrieved from the Dilated Text Convolutional Neural Network (DTCNN). After achieving two different features, it is upgraded into weighted fused features, in which the weight is tuned by an Adaptive Controlling Parameter-based Chameleon Swarm Algorithm (ACP-CSA). Finally, the fused features are fed as given to the Dilated Adaptive Deep Temporal Convolution Network with Bi-directional Long Short-Term Memory (DADTCN-Bi-LSTM) for predicting the fake news. The analysis is further performed by tuning the parameters in the model using developed ACP-CSA. The efficiency of the model is investigated and results are conducted. Thus, the analysis of the suggested system shows 95 regarding accuracy, sensitivity, and specificity. The analysis of the F1-score achieves the value of 90 in the developed model. Hence, the findings demonstrate that it achieves a better detection process to evade the existence of misinformation.

Keywords

Fake news detection visual geometry group 16 dilated text convolutional neural network adaptive controlling parameter-based chameleon swarm algorithm dilated adaptive deep temporal convolution network with bi-directional long short-term memory

1. Introduction

In this modern period, sharing information and connecting with people is a rapid process. Owing to rapid technology development, all users can share details through smartphones, and, social media become a rapid and popular platform for sharing news articles.¹ So, it becomes difficult to identify the exact beginning news spreader article.² This possibility and circulating the information easily creates the chance to increase the fake news and give the wrong impression about the ordinary people in the society.³ Fake news creation intends to appear as reliable information on the social media platform. This is evident owing to the sharing of widespread information every day without the truth of the content.⁴ The scope of sharing fake news is to manipulate the opinion of the communal for political gain and financial gain.⁵ The fake information also shows the bad effects of an investment in large infrastructure and stock price.⁶ Fake information is classified into three types: propagation, creation, and publication. The important reason for the fake news is the frauds.⁷ They quickly spread fake information with particular ideas to mislead the people, and also publish false articles in social media.⁸ It is complex to determine the value of daily content through these methods for the news shared by various frauds.⁹

Fake news creation is mainly obtained with content based on text.¹⁰ Conventional learning techniques for recognizing fake news are classified as learning on the basis of news content and social content. The news content focuses on the article news, style, malign news, etc..,¹¹ and the social context focuses on the comments, user types, shares, likes, etc. The method based on news content focuses on extracting various contents from the fake information article that includes both style and content. But, the method based on news content mainly focuses on checking the reality present in the news. Moreover, fake information spreaders have an evil plan to publish false information; they require an integration of false and true interesting information and also need the person not present in the story based on true information to make the people responsible.¹² The method based on style focuses on the written styles of the creators and influencers for publishing fake information. These fake news detection methods are not enough to detect the frauds.¹³

Nowadays, social media users are always connected to particular groups of people having similar mindsets is termed a user community.¹⁴ They are the main reason to share fake information owing to the common observation about publishing articles.¹⁵ They are groups of people with similar mindsets on social media platforms, where non-paralleled thoughts are disapproved and rejected by the way of checking the majority. For example, on Facebook, many people viewed the comments on the post provided by the user, and some liked the comment as one kind of echo chamber.¹⁶ Some people give bad comments about a person, which may make a bad impression on the particular person. Additionally, on Facebook, the default increases the number of comments regarding the likes and replies received from the friends of the user.¹⁷ Regarding the above-described challenges, we developed the effectual method.

The key components of the implemented multimodal fake detection system are elaborated below.

To propose the multimodal fake news detection model in order to classify the fake news spreaders and misinformation published by the spreader to support the society and an individual.

To establish the DTCNN and VGG-16 models are utilized to extract the text features and image features. Thus, these features are optimally selected by the developed ACP-CSA, and also weights of these features are optimized to improve accuracy.

To develop the DADTCN-Bi-LSTM model to identify fake news. These parameters are optimized in the developed DTCN-Bi-LSTM model by the developed ACP-CSA to maximize classification accuracy.

To execute the operation, the weights are tuned by the designed ACP-CSA algorithm. In the classification process, the parameters tuned by developed ACP-CSA are a count of hidden neurons in DTCN, count of hidden neurons in Bi-LSTM, epoch counts in DTCN, and epoch count in Bi-LSTM to minimize the FPR and increase the sensitivity, accuracy, and F1-score.

To estimate the detection process, the developed method is compared with different existing models to give more accuracy and precision.

The flow for the upcoming works is detailed as follows. 2^nd section gives the details about the related work. 3^rd section has the details for an intelligent model of multimodal fake news detection utilizing optimal weighted integration with dilated adaptive deep learning. 4^th section describes the heuristic-based optimal weighted integration for the multimodal fake news detection process. 5^th section describes the details of the results and discussion. Finally, the 6^th section gives the conclusion.

2. Literature survey

2.1 Related works

In 2021, Jiang et al.¹⁸ have developed three models based on deep learning and five models based on machine learning on real datasets of various scales. To determine the presentation of the model, the authors utilized metrics such as precision, accuracy, F1 score, and recall, and, also the advanced version of McNemar's testing machine was used to calculate the performance of the technique. Then, the developed stacking method attained the best accuracy value in the FDnugget and Information Security and Object Technology (ISOT) dataset. Hence, they recommended that method to detect the fake information.

In 2021, Ni et al.¹⁹ have developed the method of Multi-View Attention Networks (MVAN) based on Neural Networks (NN) for detecting fake information and helping to provide true news on social platforms. The developed MVAN method included the propagation model attention and semantic attention model. These two models helped to capture the real news, clues of the content present in the tweet, and also the fake news spreader. Moreover, these two models, based on attention, could find the keyword obtained in fake information text and doubtful users in a spreading system. The authors analyzed the model by using the real-world dataset, and also the result proved that the MVAN outperformed other models by an average accuracy of 2.5% and produced a sufficient explanation.

In 2020, Dong et al.²⁰ have recommended a method named Semi-Supervised Learning (SSL) based on deep learning. The developed model has two paths; one lane was for unsupervised learning and the other lane was for the supervised learning. The path of supervised learning had the ability to learn limited labeled data amount, but unsupervised learning learned the greatest unlabeled data amount. Moreover, the two paths of the developed models were tuned to conclude the mechanism of the implemented SSL technique. Furthermore, the authors developed the CNN for extracting the features present in the low range of unlabeled and labeled data. The CNN-based SSL method was analyzed and tested with PHEME and LIAR datasets. The result demonstrated the designed method was able to detect the fake information very effectively with the support of several numbers of labeled data.

In 2021, Song et al.²¹ have suggested the fake information recognition method regarding temporal graphs. The developed model fused the structure, temporal news, and content semantics. Particularly, the designed method could design temporal pattern evaluation of the real-world information processed regarding the continued-time dynamic diffusion network. The authors performed effective analyses of real-world datasets. The results displayed the developed technique performed better than other fake information detection methods.

In 2021, Jing et al.²² have developed Multi-model Progressive Fusion Network (MPFN) for detecting misinformation. The developed MPFN model captured the representative news of every process at various stages and achieved the fusion among the modalities at the different and same stages according to the mixture for establishing the strongest connection among every modality. Particularly, the authors utilized the transformer model, which was very effective based on computer vision approaches, as the visual features extractors progressively sample the feature at various stages and integrated each feature obtained from the text features extractors and the image domain data at various stages to well graining the model. Moreover, the authors designed the fusion task to establish the connections among the modalities, which could enhance the working performances and therefore exceed other network models. The authors conducted effective analyses on the datasets of Twitter and Weibo, and the developed model attained an accuracy of 83.3% in the Twitter dataset, which was 4.3% more than other techniques. The result proved the developed MPFN framework successfully identified the fake information, and also the developed method achieved the improved stage by using the strongest modalities fusion techniques and joining the various stages of news from every modality.

In 2021, Song et al.²³ have suggested the multimodal misinformation recognition model regarding the Cross-modal Attentions Residuals and Multichannel convolution neural Network (CARMN). The developed model effectively extracted the related information relevant to the target modalities from one source to other source modalities when continuing the distinctive news of target modalities. Multichannel Convolution neural Networks (MCN) mitigated the manipulation of the unwanted information that may created by the component of cross-modal fusion by the extraction of textual features from the fused and original textual news simultaneously. The authors conducted different analyses on real-world datasets and proved the developed framework outperformed other conventional models and learned more feature representations.

In 2021, Kaliyar et al.²⁴ have proposed a model based on social and news content with tensor. The designed framework had various filters in every layer as well as dropouts. The developed task utilized the PolitiFact and BuzzFeed datasets. The detected result proved that the developed Echo-chamber Fake news recognition based on Deep network (Echo-FakeD) task performed better than other current frameworks. The developed model has obtained 92.30% accuracy in fake information detection.

In 2020, Agarwal et al.²⁵ have constructed a fake news detection model based on a deep learning technique that predicted the article's nature. The developed model was exclusively used in the stage of text processing and was insensible to the credibility and history source of the author. The authors experimented and discussed the suggested model by utilizing the word (GloVe) for processing the text constructed the vector space word, and established the lingual relationships. The developed method had attained the greatest result in predicting fake information with the help of GloVe. Moreover, to confirm the prediction quality, different methods of the parameter were recorded and tuned to provide a better result. By adding the dropout layers between other distinctions, the overfitting problems were reduced in the developed model. The large amount of input feature was trained to provide accurate results which were more than other existing frameworks.

2.2 Problem statement

Spreading fake news widely affects the lives of the people; it damages the person's reputation and also affects the potential power. In recent days, the fast growth of social media platforms namely Twitter, Facebook, etc, spread fake news rapidly to the internet, which affects people's judgment and lives. Different models have been developed recently to identify fake news spreaders. Table 1 shows the merits and disadvantages of conventional fake news detecting techniques. The stacking model¹⁸ helps to enhance the isolated performance of the model and also it utilizes an advanced version of McNemar's to determine the accuracy. However, it has a minimum number of fake news data that is only collected for recognizing the fake news. And it has no accepted configurations. MVAN¹⁹ model requires fewer amounts of data to compute. However, it does not consider the reply of the users and it requires more time during operation. CNN-based SSL²⁰ accelerates the message verification process in a high-speed road traffic circumstance of a vehicle identifier. However, it has a small complexity during implementation, and more consideration is needed to choose the labeled ratio. Temporal Graph Network for Fake News Detection (TGNF)²¹ has the ability to handle challenging, structured data. However, TGNF has several difficulties in propagating the information and also the TGNF model can only process limited amount of tweets. MPFN²² reduces the negative infodemic effects. However, it has some complexity in addressing fake news in some languages. CARMN⁶ plays better in reducing noise information. However, the visual information is not applicable in this model. EchoFakeD²³ requires low consumption time. However, temporal information-based analyses are not performed. CNN²⁴ gives the best extracting features within a short time. However, it does not provide any ratings for creditability depict. To answer all these issues, we implemented a novel strategy for a multimodal fake news detection model.

Table 1.
Features and challenges of existing fake news detecting techniques.

Author [citation] Techniques Advantages Drawbacks

Jiang et al.¹⁸ Stacking model
It helps to increase the isolated presentation of the model.

An advanced version of the McNemar's is utilized to determine the accuracy.

Minimum numbers of fake news data are only collected for recognizing the fake news.

This has no accepted configurations.

Ni et al.¹⁹ MVAN
It requires fewer amounts of data to compute.

It does not consider the replies of the users.

It requires more time during operation.

Dong et al.²⁰ CNN-based SSL
It improves the message confirmation operation in a high-speed road traffic circumstance of a vehicle-identifier.

It has a small complexity during implementation.• More consideration is needed to choose the labeled ratio.

Song et al.²¹ Temporal Graph Network for Fake news detection (TGNF)
It has the ability to handle challenging, structured data.

It has several difficulties in propagating the information.

It can process only several tweets.

Jing et al.²² MPFN
It reduces the negative infodemic effects.

It has some complexity in addressing fake news in some languages.

Song et al.²³ CARMN
It plays better to reduce the noise information.

Visual information is not applicable in this model.

Kaliyar et al.²⁴ EchoFakeD
It requires low consumption time.

Temporal information-based analyses are not performed.

Agarwal et al.²⁵ CNN
Within a short time, it gives the best extracting features.

It does not provide any ratings for creditability depict.

Author [citation]	Techniques	Advantages	Drawbacks
Jiang et al.¹⁸	Stacking model	It helps to increase the isolated presentation of the model. An advanced version of the McNemar's is utilized to determine the accuracy.	Minimum numbers of fake news data are only collected for recognizing the fake news. This has no accepted configurations.
Ni et al.¹⁹	MVAN	It requires fewer amounts of data to compute.	It does not consider the replies of the users. It requires more time during operation.
Dong et al.²⁰	CNN-based SSL	It improves the message confirmation operation in a high-speed road traffic circumstance of a vehicle-identifier.	It has a small complexity during implementation.• More consideration is needed to choose the labeled ratio.
Song et al.²¹	Temporal Graph Network for Fake news detection (TGNF)	It has the ability to handle challenging, structured data.	It has several difficulties in propagating the information. It can process only several tweets.
Jing et al.²²	MPFN	It reduces the negative infodemic effects.	It has some complexity in addressing fake news in some languages.
Song et al.²³	CARMN	It plays better to reduce the noise information.	Visual information is not applicable in this model.
Kaliyar et al.²⁴	EchoFakeD	It requires low consumption time.	Temporal information-based analyses are not performed.
Agarwal et al.²⁵	CNN	Within a short time, it gives the best extracting features.	It does not provide any ratings for creditability depict.

3. An intelligent model of multimodal fake news detection using optimal weighted integration with dilated adaptive deep learning

3.1 Developed multimodal fake news detection model

In this generation, fake news is published through television and social media every day. The misinformation affects the reputation and financial aspects of the people. So, there is efficiency in detecting fake news creators. Several models have been suggested to categorize the wrong information publishers. However, all those models have several challenges while detecting fake news publishers. Fake news is created by the integration of real and false partial details like facts and remaining like wrong. So, the CNN model is implemented to detect fake news. But, it has several limitations and they are, it has a complexity during implementation, it requires more consideration to choose the labeled ratio, and needs more data during implementation. Thus, it is significant to identify the fake news publishers and also misinformation. For these drawbacks, we implemented a new strategy to recognize fake news regarding deep learning approaches. The graphical representation of the developed fake news detection framework is illustrated in Figure 1.

Figure 1.

Graphical illustration of the recommended fake news detection framework.

The implemented multimodal fake news detection-based deep learning system is utilized to support people and helps to deliver the facts about the news. The text data is taken from a sufficient dataset. In pre-processing, the text data is progressed as input for further processing. After finishing the procedure, it provides the pre-processed data as the output. The features are fused together with the help of pre-processed data. The VGG-16 and DTCNN techniques are used to extract the image and text features. The extracted text and image features from the DTCNN and VGG-16 models are selected optimally by using the developed ACP-CSA. Similarly, the weights from the VGG-16 and DTCNN models are tuned by utilizing the developed ACP-CSA. Then, the optimally selected features and weights are together integrated. Finally, it offers weighted fused features as the output. Then, these weighted fused features are given into the detection process. In this process, the weighted fused features are processed by the developed DADTCN-Bi-LSTM model. Here, the features are tuned using developed ACP-CSA. The parameters such as hidden neuron count in DTCN, epoch count in DTCN, hidden neuron counts in Bi-LSTM, and epoch count in Bi-LSTM are optimized for enhancing the accuracy, sensitivity, and F1-score and also for reducing the FPR. At last, it provides the classified outcomes by using the averaging process.

3.2 Fake news detection dataset

The developed multimodal fake news recognition framework gathers the sample text data and images from the Twitter dataset.

Dataset 1 (“Twitter dataset”): The required sample text data and images are taken from the Twitter dataset using the link “https://www.dropbox.com/s/40xl8dsmv4hliip/Twitter_Dataset.zip?dl = 0&file_subpath=%2FTwitter_Dataset%2Ftwitter%2Fpartition: access date: 2023-05-06”. It contains 147 image files and 147 text files.

Dataset 2 (“Fake news”): It is collected from a given link (“https://www.kaggle.com/competitions/fake-news/data?select = train.csv”). In this dataset, several attributes such as title, author, text, label, and ID are considered for training the data. For labeling the data, 0 and 1 are assumed which is considered an unreliable and reliable form.

Dataset 3 (“WELFake”): The WELFake dataset and also it is given in the following link (“https://www.kaggle.com/datasets/saurabhshahane/fake-news-classification”). In total, 72,134 news articles are presented, where the real and fake news consists of 35,028 and 37,106. It contains the serial number, title, text, and label as attributes in this dataset. The sample text data and images are represented as $s a_{r}^{T}$ and $s a_{p}^{I}$ .

3.3 Proposed ACP-CSA

The developed ACP-CSA algorithm is used for optimization. The parameters optimized by the developed ACP-CSA are hidden neuron count in DTCN; epoch count in DTCN, hidden neuron count in Bi-LSTM, and epoch count in Bi-LSTM for enhancing the accuracy, sensitivity, F1-score and also for reducing the FPR. The CSA algorithm is chosen for having major benefits; it is very effective in solving optimization problems, enhances working ability, and also has generalization and flexibility. But, it has the following drawbacks; implementation quality is low and it requires more memory for updating the velocity. To address these limitations, we developed the ACP-CSA algorithm for tuning. The adaptive concept is introduced in determining different random parameters of CSA using expressions equations (1)–(3).

\begin{aligned} α & = 4.0 \end{aligned}

(1)

\begin{aligned} β & = c e i l (\frac{B}{α}) \end{aligned}

(2)

\begin{aligned} γ & = f l o o r (\frac{B}{α}) \end{aligned}

(3)

The variable

β = c e i l (\frac{B}{α})

is denoted as the population of CSA. The term

β

γ

and

α

are the controlling parameter in ACP-CSA.

CSA²⁶: The following sections describe the characteristics of a chameleon while pasturing.

Initialization and function evaluation: It works to initialize the tuning process. The population of the chameleons is represented by B. Equation (4) is used to determine the chameleon's position.

z_{u}^{j} = [z_{u, 1}^{j}, z_{u, 2}^{j}, \dots z_{u, e}^{j}]

(4)

The term

j = 1, 2.. u

denotes the number of current iterations, e specifies the dimension and the term denotes

j^{t h}

chameleon position.

The original population of chameleons is generated based on the total count of the chameleons is elaborated in equation (5).

z^{j} = m_{k} + s \times (v_{k} - m_{k})

(5)

The variable

z^{j}

denotes the original vectors of

j^{t h}

a chameleon, and a random number is taken as s at

[o, 1]

range.

Search for prey: The chameleon's behavior during pasturing is determined using equation (6).

z_{u + 1}^{j, k} = {\begin{array}{lc} z_{u}^{j, k} + q_{1} (O_{u}^{j, k} - F_{u}^{k}) s_{2} + q_{2} (F_{u}^{k} - z_{u}^{j, k}) s_{1} & s_{j} \geq O_{q} \\ z_{u}^{j, k} + μ ((v^{k} - m_{k}) s_{3} + m_{c}^{k}) s g (r n d - 0.5) & s_{j} \geq O_{q} \end{array}

(6)

In equation (6), the term

z_{u + 1}^{j, k}

denotes the chameleon's new location. The term

μ

denotes the random parameter. And

z_{u}^{j, k}

denotes the chameleon's current area,

O_{u}^{j, k}

denotes the chameleon's best-scored position, and

F_{u}^{k}

describes the chameleon's universal best location. The terms

q_{1}

q_{2}

describe the positive number. The terms,

s_{1}

s_{2}

and

s_{3}

represent the random number. The term

s_{j}

expresses the arbitrarily developed random number at the index. Moreover, the term

O_{q}

describes the discerned prey of the chameleon. The connection form of O and F is determined using equation (7).

Q (s_{1}) = s_{1} O + (1 - s_{1}) F

(7)

The random value categorizes the line paths that are described in equation (8).

P (s_{1}, s_{2}) = s_{2} R + (1 - s_{2}) Q 0 \leq s_{2} \leq 1

(8)

Equations (7) and (8) describe chameleons as having a high potential for exploring every location. The features in these two equations are derived by conflict with the important features from equation (6), where

F_{u}^{k} = F

z_{u}^{j, k} = Q

u_{u + 1}^{j, k} = P (s_{1}, s_{2})

and

O_{u}^{j, k} = O

. The condition

O_{q} < 0.1

is developed to let the chameleons explore the arbitrary position in a search region to improve the searching capabilities of the CSA. This function is elaborated in equation (9).

μ = γ {e^{(\frac{- α u}{S})}}^{β}

(9)

In equation (9), the terms,

β

γ

and

α

are the constants in the existing CSA algorithm that are utilized for controlling the exploitation and exploration capabilities, and these are updated by adaptive concepts as equations (1)–(3).

Chameleons’ eye rotation: To recognize the locality of the prey, the chameleons rotate their eyes. Chameleons have the capacity to rotate their eyes at 360. This rotation is mathematically modeled for realizing the behavior of hunting. The chameleons move and rotate their eyes according to the prey's position. Hence, the location of the chameleons can be upgraded from the rotation of the eyes as explained in equation (10).

z_{u + 1}^{j} = z s_{u}^{j} + z_{u}^{j}

(10)

Here, the term

z_{u + 1}^{j}

denotes the new position of the chameleon later than the eye turning

z_{u}^{j}

and represents the current locality of the chameleon before eye rotation. The term

z s_{u}^{j}

describes the chameleon's rotating coordinate inside the searching Space. The eye rotation of the chameleon along the y and x direction is explained in equations (11) and (12).

Q_{y} = [\begin{array}{ccc} 1 & 0 & 0 \\ 0 & \cos θ & - \sin θ \\ 0 & \sin θ & \cos θ \end{array}]

(11)

In equation (11), the term

θ

denotes the chameleons rotate their eyes at the x-axis.

Q_{z} = [\begin{array}{ccc} \cos ϕ & 0 & \sin ϕ \\ 0 & 1 & 0 \\ - \sin ϕ & 1 & \cos ϕ \end{array}]

(12)

In equation (12), the term

ϕ

denotes the chameleons rotate their eyes at the x-axis.

Hunting prey: Chameleons hunt their prey while the prey is near to the chameleons. Suppose the chameleons get the best prey, that's it optimal position. They utilize their tongue for attaching the prey. Therefore, the position is upgraded as it straightens its tongue fully to attack. The chameleon's tongue velocity is modeled based on equation (13).

w_{u + 1}^{j, k} = ω w_{u}^{j, k} + d_{1} (F_{u}^{k} - z_{u}^{j, k}) s_{1} + d_{2} (O_{u}^{j, k} - z_{u}^{j, k}) s_{2}

(13)

Here, the term

w_{u + 1}^{j, k}

denotes the chameleon's new velocity, and

z_{u}^{j, k}

describes the chameleon's current location. The terms

d_{2}

and

d_{2}

represents the constant value that control

F_{u}^{k}

and

O_{u}^{j, k}

tongue of the chameleon. And, the term

ω

defines the weight of inertia. The weight of inertia is reduced by the generations iterated as described in equation (14).

ω = {(1 - \frac{u}{S})}^{(ρ \sqrt{\frac{u}{S}})}

(14)

Here, the term

ρ

indicates the positive constant for managing the capacity of exploitation. The chameleon's location after the projection of the tongue towards the prey is detailed in equation (15).

z_{u + 1}^{j, k} = z_{u}^{j, k} + (\frac{{(w_{u}^{j, k})}^{2} - {(w_{u - 1}^{j, k})}^{2}}{2 b})

(15)

In equation (15), the term

w_{u - 1}^{j, k}

denotes the chameleon's preceding speed and b describes acceleration rates for tongue projections of the chameleon. All these expressions are useful for exploring various random positions of the chameleons and defining the best location to find the optimal prey. The pseudocode for the designed ACP-CSA is detailed in Algorithm 1 and the flowchart for the recommended ACP-CSA is described in Figure 2.

Figure 2.

Flowchart of the proposed ACP-CSA.

4. Heuristic-based optimal weighted integration for multimodal fake news detection model

4.1 Pre-processing of data

The text data $s a_{r}^{T}$ is put as the input in the pre-processing method. It converts the data into the understanding format. The pre-processing method helps to extract the unwanted words from the sentences. Here, the text data get pre-processed by using 1) punctuation and special character removal, and 2) stemming is briefly discussed as follows.

Punctuation and stop word removal: This helps to remove conjunctions, adverbs, etc. It used to focus more on the main content by reducing the low-level cases.

Stemming: It is utilized to minimize the meaningless words in the sentence.

After completing these stages, it provides the pre-processed text data as the output and it is represented as $p r e_{r}^{p r}$ .

4.2 VGG16-based feature extraction

VGG-16²⁷ model is used to extract the image features and it is described by $s a_{p}^{I}$ . It is the CNN-based model. It is the improved model of AlexNet model by swapping the largest size of the kernel filter by multiple the kernel size as $3 \times 3$ , which is placed one after the another layer. The first two layers have 64 features of kernel filters and $3 \times 3$ sizes. The input images are carried via the first two convolutional layers. The output images are carried to maxpolling layers with 2 strides. The next two layers (3 and 4) have 124 numbers of features of kernel filters, and the size of the filter is $3 \times 3$ . These two layers followed the maxpolling layers with 2 strides. The resulting image output is decreased to the size of $56 \times 56 \times 128$ . The next three layers (5 to 7) are convolutional layers and have 256 features of kernel filters and $3 \times 3$ sizes. This fifth to seventh layers followed the maxpolling layers with 2 strides. The next four layers (7 to 11) are a set of two convolutional layers and have 512 features of kernel filters and have $3 \times 3$ sizes. These seventh to eleventh layers followed the maximum pooling layers with 1 stride. The last two layers (14 and 15) are the hidden layers which are fully connected and having units of 4096 following the softmax output layers with 1000 units. The extracted image features are expressed by $E x_{q}^{I}$ .

4.3 DTCNN-based feature extraction

DTCNN²⁸ architecture is utilized to extract the text features and alternate the input feature $p r e_{r}^{p r}$ of the model, utilizing 2D convolution as an alternative to 3D convolution. The first and second layers have 32 numbers of filters $1 \times 3$ and $1 \times 2$ a size of kernel. The third and fourth layers have 32 numbers of filters, $1 \times 2$ and $2 \times 1$ . These four layers are fed and concatenated to the layers which are fully connected and contain 64 numbers of neurons to perform the classification process.

Convolution layer: This layer is applied as the input for creating the feature map and getting classified features. Let's take $E (n, m)$ as the input feature and of size $(h, i)$ , the term $c (h, i)$ denotes the biases. The output for this layer is shown in equation (16).

Q (n, m) = g_{x}^{(h \times i)} \otimes E (n, m) + c^{(h, i)}

(16)

ReLU Activation layer: The ReLU activations layers present after each convolution layer are used for normalizing the output feature. This ReLU layer helps for learning complex tasks to reduce the vanishing gradient possibilities and reduce the cost of computation. The function of the ReLU activation layer is determined by using equation (17).

g (γ) = y (0, γ)

(17)

Concatenation layer: This layer is used to integrate the different input features in a continuous way.

Fully connected layer: This fully concatenated layer has multiple layers, which are connected to the preceding layer. The activation functions of the neurons are determined by using the multiplicative matrix weight added by the value of the offset. Let's take an input as $α$ with the size of v, and s denotes the neuron's number. An output feature of the fully connected layer is identified using equation (18).

d (α) = υ (T * α)

(18)

In equation (18), the attribute

υ

indicates the activation functions.

Dropout layer: This layer arbitrarily deactivates or activates the extrovert hidden unit at every training phase's output. It is also used to lessen the overfitting problem.

Classification layer: This classification layer is the last layer, which works for the process of classification regarding the extracted features by every preceding layer. This is the modern layer of ANN with the activation functions of sigmoid or softmax.

Loss function: For the binary-class classification task of text, DTCNN is employed to minimize the binary cross entropy using a sigmoid function, is explained in equation (19). The DTCNN is employed to minimize the categorical cross entropy by using the softmax function for multi-class recognition process is explained in equation (20).

\begin{aligned} b i & = - \frac{1}{n} \sum_{j}^{n} \sum_{k}^{d} z_{j, k} \log (σ (u_{j, k})) - (1 - z_{j, k}) \log (1 - σ (u_{j, k})) \end{aligned}

(19)

\begin{aligned} b i & = - \frac{1}{n} \sum_{j}^{n} \sum_{k}^{d} z_{j, k} \log (\frac{e^{u_{_{j, k}}}}{\sum_{s = 1}^{d} e^{u_{_{j, k}}}}) \end{aligned}

(20)

Here, the term j denotes the training instance's index, k represents the class's index,

u_{j, k}

denotes the fully connected layer's output, and

z_{j, k}

describes the actual values of

j^{t h}

the sample and

k^{t h}

class. The extracted text features are denoted by

E x_{s}^{T}

4.4 Weighted feature fusion with developed ACP-CSA

The image features $E x_{q}^{I}$ are extracted using VGG-16 and the text features $E x_{s}^{T}$ are obtained from the DTCNN model that are considered in this weighted feature fusion phase. The image and text features from the VGG-16 framework and DTCNN models are initially used to select optimal features by the designed ACP-CSA. The optimally selected features from image and text data are indicated by ${Im}_{b}^{o p t}$ and $T e_{g}^{o p t}$ . Similarly, two different weights are optimized with the support of the suggested ACP-CSA, and the weights are described as $W_{b}^{1}$ and $W_{g}^{2}$ . Then, the optimally selected features and weights are integrated to give the weighted fused features. The mechanism of the weighted fused feature is detailed in equation (21).

f u_{p}^{w t} = W_{b}^{1} * {Im}_{b}^{o p t} + W_{g}^{2} * T e_{g}^{o p t}

(21)

Here, the term

W_{b}^{1}

and

W_{g}^{2}

denotes the weights integrated with image and text features, the terms

{Im}_{b}^{o p t}

and

T e_{g}^{o p t}

denote the optimally selected image and text features from VGG-16 and DTCNN. The weighted fused features output is denoted by

f u_{p}^{w t}

. The fitness function of the developed concept is explained in equation (22).

S_{1} = \underset{{W_{b}^{1}, {Im}_{b}^{o p t}, W_{g}^{2}, T e_{g}^{o p t}}}{\arg min} (\frac{1}{C + V})

(22)

In equation (22) the variable

S_{1}

is defined as the fitness, C denotes the correlation coefficient, and V denotes the variance. The weights

W_{b}^{1}

and

W_{g}^{2}

are in the range of

[0 .01 - 0 .99]

. The functions of C and V are described in equations (23) and (24).

\begin{aligned} V & = \sum_{p = 1}^{t} {(f u_{p}^{w t} - \bar{f u_{p}^{w t}})}^{2} \end{aligned}

(23)

\begin{aligned} C & = \frac{R_{a b}}{R_{a} R_{b}} \end{aligned}

(24)

In equations (23) and (24), the variable

f u_{p}^{w t}

describes the weighted fused futures and

\bar{f u_{p}^{w t}}

describes the mean value of the weighted fused futures. The term

R_{a b}

is represented as sample covariance and

R_{a} R_{b}

describes sample standard deviation. The diagrammatic explanation for the developed model is detailed in Figure 3.

Figure 3.

Diagrammatic representation of weighted feature fusion with developed ACP-CSA.

5. Efficient multimodal fake news detection model with dilated adaptive deep temporal convolution network with bi-directional long short-term memory

5.1 DTCN

DTCN²⁹ is based on the group of RNNs, and every layer has a 1-D block with dilation. The dilation feature is increased to ensure the appropriate large context to provide the advantages of the speech signals. The waveNet with dilation feature has great success in audio-generating tasks. Stacking dilation convolutions gives a huge receptive field with several layers because the dilation factor increased exponentially. This permits the TCN to take the temporal dependencies of several resolutions with an input series. The TCN develops the temporal chain of command; the top layers can permit the series of long inputs and can learn the representations on a larger scale. The local details from the below layers are propagated via residual connections.

The TCN has two key elements. 1) dilated convolution and 2) residual connection. In the dilated convolution, the following mechanism is described in equation (25).

(y * e^{l}) (q) = \sum_{t + e u = q} y (t) l {(u)}^{'}

(25)

Here, the term y denotes the input signals, l determines the filters and, the term e denotes the dilator feature. The receptive fields are increased by increasing the dilation factors e. The DADTCN in this model has 4 convolutional layers with the dilator factor. In Figure 4, each 1D convolution element is the residual block that has the dilated convolutions with one layer, two-layer convolution with

1 \times 1

size, two numbers of PReLU, and also has two numbers of normalizations layers. The global normalizations are applied to convolution filters during the normalizing process. The pictorial representation of the DTCN technique is displayed in Figure 4.

Figure 4.

A pictorial illustration of the DADTCN framework.

5.2 Bi-LSTM

Bi-LSTM³⁰ is the series processing technique. It has two LSTMs, the first LSTM taking the text input and the second LSTM processing with a rearward copy in the series. The Bi-LSTM framework has the advantages of the LSTM and CNN model. It has two sections, and they are backward and forward Bi-LSTM. The formulation of Bi-LSTM is provided in equations (26) and (27)

\begin{aligned} i_{u} & = N_{u} . \tanh (g_{u} . B_{u - 1} + j_{u} . \bar{B_{u}}) \end{aligned}

(26)

\begin{aligned} {i_{u}}^{'} & = {N^{'}}_{u} . \tanh ({g^{'}}_{u} . {B^{'}}_{u - 1} + {j^{'}}_{u} . \bar{{B^{'}}_{u}}) \end{aligned}

(27)

The equation (26)

i_{u}

is the forward Bi-LSTM's output. The term u stands for the time. The term

i_{u}'

signifies the backward Bi-LSTM's output. The term

\bar{B_{u}}

denotes a new vector that is developed using tanH function. The term

g_{u}

represents the forget gate,

B_{u - 1}

the state of cellular, and

j_{u}

denotes the input gate. The output of the Bi-LSTM model is described in equation (28).

z_{u} = [i_{u}, i_{u}^{'}]

(28)

Here, the term

z_{u}

describes the output of Bi-LSTM. Thus, it gathers the data from rearward and onward directions. The diagram for the Bi-LSTM model is detailed in Figure 5.

Figure 5.

A pictorial illustration of the Bi-LSTM framework.

5.3 Implemented DADTCN-Bi-LSTM-based multimodal fake news detection model

The weighted fused feature $f u_{p}^{w t}$ is passed through the input in the classification operation. In this implemented DADTCN-Bi-LSTM technique, the parameters are optimized by the implemented ACP-CSA algorithm to give the best-classified solution. The parameters tuned in the DADTCN-Bi-LSTM model are hidden neuron count in DTCN, epoch count in DTCN, hidden neuron count in Bi-LSTM, and epoch count in Bi-LSTM for enhancing the accuracy, sensitivity, F1-score, and also for reducing the FPR. After finishing this process, the weighted fused features provide the classified outcome of detected fake news as the output. The fitness operation for the classification process is explained in equation (29).

S_{2} = \underset{{h_{n}^{D T}, e_{m}^{D T}, {h_{w}}^{L S}, {e_{e}}^{L S}}}{\arg min} (\frac{1}{A + S y + F r} + F 1)

(29)

In equation (29), the terms A,

S y

F r

and

F 1

denote the accuracy, sensitivity, F1-score, and FPR respectively. The variable

S_{1}

declares the fitness function in the classification process. Moreover, the terms

h_{n}^{D T}

and

{h_{w}}^{L S}

describe the count of hidden neurons in DTCN in the range of

[5, 255]

and the count of hidden neuron Bi-LSTM in the range of

[5, 255]

. Similarly,

e_{m}^{D T}

and

{e_{e}}^{L S}

denotes epochs in DTCN in the ranges of

[5, 50]

and epochs in Bi-LSTM in the range of

[5, 50]

. The operations of A,

S y

F r

and

F 1

are described in the below equations.

\begin{aligned} A & = \frac{(p + q)}{(p + q + r + s)} \end{aligned}

(30)

\begin{aligned} S y & = \frac{p}{p + s} \end{aligned}

(31)

\begin{aligned} F r & = \frac{r}{s + r} \end{aligned}

(32)

\begin{aligned} F 1 & = \frac{2 * p}{2 * (p + r + s)} \end{aligned}

(33)

Here, the variables q, p, s and r are represented as the true negative, true positive, false negative, and false positive. The pictorial illustration of the designed DADTCN-Bi-LSTM model is described in Figure 6.

Figure 6.

Pictorial representation of DADTCN-Bi-LSTM-based fake news detection model.

6. Results

6.1 Simulation setting

The designed fake news detection process was executed using Python. The implemented system considered the number of populations as 10, and also the maximum iteration was 50. The developed fake news model was compared and analyzed with the classifiers, namely, Autoencoder,³¹ DTCN,²⁹ Bi-LSTM,³⁰ and DTCN-Bi-LSTM.³¹ The algorithms like Dingo Optimizer (DO),³² Eurasian Oystercatcher Optimizer (EOO),³³ Reptile Search Algorithm (RSA),³⁴ and Chameleon Swarm Algorithm (CSA)²⁶ were considered for the validation.

6.2 Evaluation metric

The following performance measures are considered to analyze the multimodal fake news identification system.

(a)
Precision: It shows the relevant data from the obtained instance which is determined using equation (34).
$P = \frac{p}{p + r}$
(34)
(b)
Negative Predictive Value (NPV): The NPV $N p$ is explained as the ratio of detecting the actual test outcomes that have been negatively detected while testing the data. It is evaluated using equation (35).
$N p = \frac{q}{s + q}$
(35)
(c)
False Discovery Rate (FDR): FDR measure $F r$ is determined as the number of rejected data that is obtained from the positive rates of true and false which are calculated in equation (36).
$F r = \frac{r}{r + p}$
(36)
(d)
Specificity: Specificity $S y$ is defined as the proportion of the negative values that need to be properly identified and also it is formulated in equation (37).
$S y = \frac{q}{q + r}$
(37)
(e)
False Negative Rate (FNR): FNR $F n$ is measured as the positive data that is yielded from the negative test solution during the testing phase. It is measured in equation (38).
$F n = \frac{r}{q + r}$
(38)
(f)
Mathew's Correlation Coefficient (MCC): MCC $M c$ is defined to test the data with accurate outcomes identified using equation (39).
$M c = \frac{p \times q - r \times s}{\sqrt{(p + q) (p + s) (q + p) (q + s)}}$
(39)

6.3 Analysis of cost function using the developed model

The cost function of the developed fake news classification framework is explained in Figure 7. In Figure 7, the implemented ACP-CSA-DTCN-Bi-LSTM model is 1.48% more than DO-DTCN-Bi-LSTM, 0.99% more than EOO-DTCN-Bi-LSTM, 1.32% enhanced than RSA-DTCN-Bi-LSTM and 2% improved than CSA-DTCN-Bi-LSTM when the maximum iteration at 10. This proved that the efficiency of the implemented method offers enhanced performance.

Figure 7.

Cost function of the developed model.

6.4 Performance estimation of the developed model

The confusion matrix analysis of the implemented fake news model is described in Figure 8. Figure 8 shows the predicted and actual values. The implemented ACP-CSA-DTCN-Bi-LSTM model gives an accuracy value of 97.18% which is better than other existing fake news detecting techniques.

Figure 8.

Confusion matrix of the developed model.

6.5 Classifier-based analysis of implemented fake news detection model

The classifier-aided analysis of the implemented fake news recognition process is illustrated in Figure 9. In Figure 9(d), the FNR of the implemented ACP-CSA-DTCN-Bi-LSTM model is 60% improved than Autoencoder, 53.33% more than DTCN, 53.33% higher than Bi-LSTM and 35.38% more than DTCN-Bi-LSTM when the activation function is observed as ReLU. This proved the developed ACP-CSA-DTCN-Bi-LSTM model offers sufficient performance.

Figure 9.

Classifier-based analysis of the implemented fake news detection model regarding “(a) accuracy (b) F1-score (c) FDR (d) FPR (e) FNR (f) MCC (g) NPV (h) precision (i) sensitivity and (j) specificity”.

6.6 Algorithmic-based analysis of developed fake news detection process

The algorithmic-based analysis of the developed fake news detection process is detailed in Figure 10. In Figure 10(b), the F1-score of the proposed ACP-CSA-DTCN-Bi-LSTM framework is 3.95% improved than DO-ACP-CSA-DTCN-Bi-LSTM, 3.48% enhanced than EOO- ACP-CSA-DTCN-Bi-LSTM, 3.37% more than RSA-ACP-CSA-DTCN-Bi-LSTM and 2.79% higher than CSA- ACP-CSA-DTCN-Bi-LSTM when activation function is TanH. Therefore, this proved the presentation of the designed ACP-CSA-DTCN-Bi-LSTM model is much better than other algorithms.

Figure 10.

Algorithmic-based analysis of the proposed fake news detection process in terms of “(a) accuracy (b) F1-score (c) FDR (d) FPR (e) FNR (f) MCC (g) NPV (h) precision (i) sensitivity and (j) specificity”.

6.7 Accuracy-based analysis of developed fake news detection model

The accuracy-based classifier analysis of the implemented fake news detection framework is detailed in Figure 11. The weighted fused feature accuracy of the designed ACP-CSA-DTCN-Bi-LSTM model is 16.47% more than Autoencoder, 15.11% enhanced than DTCN, 13.79% superior than Bi-LSTM and 12.5% more than DTCN-Bi-LSTM (without optimization). Hence, it showed the developed fake news recognition framework has a significant role in fusing the features. The accuracy-based algorithmic analysis of the implemented fake news recognition process is shown in Figure 11. In Figure 11, the weighted fused feature accuracy of the implemented ACP-CSA-DTCN-Bi-LSTM model is 6.52% more than DO-DTCN-Bi-LSTM, 6.90% improved than EOO-DTCN-Bi-LSTM, 4.25% enhanced than RSA-DTCN-Bi-LSTM and 5.37% more than CSA-DTCN-Bi-LSTM. This ensured that the implemented process has more accuracy while fusing the features.

Figure 11.

Accuracy-based analysis of the developed fake news detection process in terms of (a) classifiers and (b) algorithms.

6.8 ROC-based analysis of implemented fake news detection process

The ROC-based analysis of the developed fake news detection process is illustrated in Figure 12. Generally, the ROC curve is determined based on the distinct threshold values of the false positive rate as well as the true positive rate. It demonstrates to show the accurate classification in the fake news process. Hence, the Area under the ROC Curve (AUC) determines the measurement of two dimensional underneath area which holds the entire ROC curve therefore it lies in the interval of 0 to 1. This showed the developed fake news performs more than other existing methods.

Figure 12.

Performance analysis based on ROC using the developed fake news detection framework.

ROC curve: The ROC curve is defined as the true relationship of sensitivity and specificity is provided for each possible cut-off value. The formula for sensitivity is given in equation (31) and also the formula for specificity is shown in equation (32).

6.9 Overall algorithm-based analysis of implemented fake news detection process

The overall classifier investigation of the developed fake news detection process is detailed in Table 2. The overall performance of the developed model is evaluated with diverse conventional methods while considering with standard evaluation metrics. Nowadays, fake news arises in a tremendous manner that's why the implemented process is performed to assist to provide accurate performance in the fake news detection framework. In Table 2, the specificity of the developed ACP-CSA-DTCN-Bi-LSTM method is, 1.80% increased than DO-DTCN-Bi-LSTM, 1.58% more than EOO-DTCN-Bi-LSTM, 1.58% enhanced than RSA-DTCN-Bi-LSTM and 1.12% higher than CSA-DTCN-Bi-LSTM. This showed the developed fake news model has more presentation metrics. In this Table analysis, the implemented process achieves the value of 95 while considering the evaluation measures such as accuracy, sensitivity, and specificity. Due to the increased accuracy value, the error gets reduced while detecting the fake news. Also, the DO-DTCN-Bi-LSTM model attains the value of 18.301 in FDR analysis. The higher error analysis seems to lower the performance of the framework. Hence, the recommended technique proved that it has attained efficient and outsourced performance when compared with the existing approaches.

Table 2.
Overall algorithm-based analysis of the developed fake news detection process.

Terms DO-DTCN-BI-LSTM³² EOO-DTCN-BI-LSTM³³ RSA-DTCN-BI-LSTM³⁴ CSA-DTCN-BI-LSTM²⁶ ACP-CSA-DTCN-BI-LSTM

Dataset 1

Accuracy 94.030 94.196 94.196 94.693 95.688

Sensitivity 93.985 93.985 93.985 94.737 95.489

Specificity 94.043 94.255 94.255 94.681 95.745

Precision 81.699 82.237 82.237 83.444 86.395

FPR 5.957 5.745 5.745 5.319 4.255

FNR 6.015 6.015 6.015 5.263 4.511

NPV 98.222 98.226 98.226 98.451 98.684

FDR 18.301 17.763 17.763 16.556 13.605

F1-Score 87.413 87.719 87.719 88.732 90.714

MCC 0.839 0.843 0.843 0.856 0.881

Terms	DO-DTCN-BI-LSTM³²	EOO-DTCN-BI-LSTM³³	RSA-DTCN-BI-LSTM³⁴	CSA-DTCN-BI-LSTM²⁶	ACP-CSA-DTCN-BI-LSTM
Dataset 1
Accuracy	94.030	94.196	94.196	94.693	95.688
Sensitivity	93.985	93.985	93.985	94.737	95.489
Specificity	94.043	94.255	94.255	94.681	95.745
Precision	81.699	82.237	82.237	83.444	86.395
FPR	5.957	5.745	5.745	5.319	4.255
FNR	6.015	6.015	6.015	5.263	4.511
NPV	98.222	98.226	98.226	98.451	98.684
FDR	18.301	17.763	17.763	16.556	13.605
F1-Score	87.413	87.719	87.719	88.732	90.714
MCC	0.839	0.843	0.843	0.856	0.881

6.10 Overall classifier-based analysis on developed fake news detection process

Table 3 illustrates the algorithmic-based analysis of the developed fake news detection model. The analysis is performed with existing techniques where the developed model is suggested to outperform different analyses. Here, the analysis is conducted with positive and negative measures using the developed system. In Table 3, the precision of the developed ACP-CSA-DTCN-Bi-LSTM framework is 20.80% increased to Autoencoder, 15.65% superior to DTCN, 15.69% improved than Bi-LSTM and 7.82% more than DTCN-Bi-LSTM. Accurate fake news is detected where the accuracy of the framework needs to be enhanced. Based on this analysis, the developed ACP-CSA-DTCN-BiLSTM model shows accurate performance. This overall algorithmic-based analysis proved the developed fake news model performs better in all metrics.

Table 3.
Overall classifier-based analysis of the implemented fake news detection process.

Terms Autoencoder³⁵ DTCN²⁹ BI-LSTM³⁰ DTCN-BI-LSTM³¹ ACP-CSA-DTCN-BI-LSTM

Dataset 1

Accuracy 89.718 91.542 91.211 93.532 95.688

Sensitivity 88.722 93.233 90.977 93.985 95.489

Specificity 90.000 91.064 91.277 93.404 95.745

Precision 71.515 74.699 74.691 80.128 86.395

FPR 10.000 8.936 8.723 6.596 4.255

FNR 11.278 6.767 9.023 6.015 4.511

NPV 96.575 97.941 97.279 98.210 98.684

FDR 28.485 25.301 25.309 19.872 13.605

F1-Score 79.195 82.943 82.034 86.505 90.714

MCC 0.732 0.783 0.769 0.827 0.881

Terms	Autoencoder³⁵	DTCN²⁹	BI-LSTM³⁰	DTCN-BI-LSTM³¹	ACP-CSA-DTCN-BI-LSTM
Dataset 1
Accuracy	89.718	91.542	91.211	93.532	95.688
Sensitivity	88.722	93.233	90.977	93.985	95.489
Specificity	90.000	91.064	91.277	93.404	95.745
Precision	71.515	74.699	74.691	80.128	86.395
FPR	10.000	8.936	8.723	6.596	4.255
FNR	11.278	6.767	9.023	6.015	4.511
NPV	96.575	97.941	97.279	98.210	98.684
FDR	28.485	25.301	25.309	19.872	13.605
F1-Score	79.195	82.943	82.034	86.505	90.714
MCC	0.732	0.783	0.769	0.827	0.881

6.11 Statistical analysis of implemented fake news detection process

The statistical analysis of the developed fake news detection process is shown in Table 4. The standard evaluation is performed where the developed model is performed with best, worst, mean, median, and standard deviation. Focusing on statistical analysis supports to rectify the problems of gradient vanishing and overfitting issues. This analysis is helpful when the developed model shows better and more accurate performance in the fake news detection process. In Table 4, the mean value of the suggested ACP-CSA-DTCN-Bi-LSTM model is 1.49% increased than DO-DTCN-Bi-LSTM, 1.18% more than EOO-DTCN-Bi-LSTM, 1.19% enhanced than RSA-DTCN-Bi-LSTM and 0.39% higher than CSA-DTCN-Bi-LSTM. This proved that the developed fake news detection model performs better in standard deviation, median, best, mean, and worst.

Table 4.
Statistical analysis of the implemented fake news detection process.

Algorithms DO-DTCN-BI-LSTM³² EOO-DTCN-BI-LSTM³³ RSA-DTCN-BI-LSTM³⁴ CSA-DTCN-BI-LSTM²⁶ ACP-CSA-DTCN-BI-LSTM

Best 1.000 1.007 1.000 1.011 1.000

Worst 1.017 1.030 1.030 1.084 1.098

Mean 1.006 1.009 1.001 1.025 1.021

Median 1.000 1.007 1.000 1.012 1.015

Standard deviation 0.807 0.447 0.584 2.176 2.552

Algorithms	DO-DTCN-BI-LSTM³²	EOO-DTCN-BI-LSTM³³	RSA-DTCN-BI-LSTM³⁴	CSA-DTCN-BI-LSTM²⁶	ACP-CSA-DTCN-BI-LSTM
Best	1.000	1.007	1.000	1.011	1.000
Worst	1.017	1.030	1.030	1.084	1.098
Mean	1.006	1.009	1.001	1.025	1.021
Median	1.000	1.007	1.000	1.012	1.015
Standard deviation	0.807	0.447	0.584	2.176	2.552

6.12 Experimental analysis of the implemented process with two experimental datasets

The experimental analysis is contrasted and validated with two experimental datasets for the developed model and also it is shown in Table 5. Also, the implemented process is also compared with diverse classifiers in terms of new datasets which is listed in Table 6. Based on this evaluation, the developed model shows 5.09%, 3.43%, 2.67%, and 1.59% improvements over DO-DTCN-Bi-LSTM, EOO-DTCN-Bi-LSTM, RSA-DTCN-Bi-LSTM, and CSA-DTCN-Bi-LSTM regarding precision metric. It seems that the accuracy of the implemented process is 97% which has the ability to solve complex problems like overfitting. When considering Table 6, the FDR rate in Autoencoder shows the value of 6.74 which is assumed to contain the higher error rate. Hence, the Autoencoder model is not efficient for detecting fake news. While analyzing these results, the developed model proved the accurate outcomes in the fake news detection framework.

Table 5.
Experimental analysis of the implemented model with new datasets using diverse algorithms.

Terms DO-DTCN-BI-LSTM³² EOO-DTCN-BI-LSTM³³ RSA-DTCN-BI-LSTM³⁴ CSA-DTCN-BI-LSTM²⁶ ACP-CSA-DTCN-BI-LSTM

Dataset 2

FPR 7.58 6.02 5.33 4.33 2.79

FNR 7.12 6.39 5.59 4.48 2.90

Accuracy 92.65 93.79 94.54 95.59 97.15

Precision 92.51 93.99 94.69 95.69 97.22

NPV 92.80 93.59 94.39 95.49 97.08

MCC 85.31 87.59 89.08 91.19 94.31

Sensitivity 92.88 93.61 94.41 95.52 97.10

Specificity 92.42 93.98 94.67 95.67 97.21

FDR 7.49 6.01 5.31 4.31 2.78

F1-Score 92.69 93.80 94.55 95.60 97.16

Dataset 3

Accuracy 93.73 94.63 95.50 96.06 97.50

Sensitivity 93.76 94.63 95.64 95.98 97.66

Specificity 93.71 94.62 95.35 96.14 97.34

Precision 93.69 94.60 95.35 96.12 97.34

FPR 6.29 5.38 4.65 3.86 2.66

FNR 6.24 5.37 4.36 4.02 2.34

NPV 93.78 94.65 95.65 96.00 97.66

FDR 6.31 5.40 4.65 3.88 2.66

F1-Score 93.72 94.62 95.49 96.05 97.50

MCC 87.47 89.25 90.99 92.12 95.00

Terms	DO-DTCN-BI-LSTM³²	EOO-DTCN-BI-LSTM³³	RSA-DTCN-BI-LSTM³⁴	CSA-DTCN-BI-LSTM²⁶	ACP-CSA-DTCN-BI-LSTM
Dataset 2
FPR	7.58	6.02	5.33	4.33	2.79
FNR	7.12	6.39	5.59	4.48	2.90
Accuracy	92.65	93.79	94.54	95.59	97.15
Precision	92.51	93.99	94.69	95.69	97.22
NPV	92.80	93.59	94.39	95.49	97.08
MCC	85.31	87.59	89.08	91.19	94.31
Sensitivity	92.88	93.61	94.41	95.52	97.10
Specificity	92.42	93.98	94.67	95.67	97.21
FDR	7.49	6.01	5.31	4.31	2.78
F1-Score	92.69	93.80	94.55	95.60	97.16
Dataset 3
Accuracy	93.73	94.63	95.50	96.06	97.50
Sensitivity	93.76	94.63	95.64	95.98	97.66
Specificity	93.71	94.62	95.35	96.14	97.34
Precision	93.69	94.60	95.35	96.12	97.34
FPR	6.29	5.38	4.65	3.86	2.66
FNR	6.24	5.37	4.36	4.02	2.34
NPV	93.78	94.65	95.65	96.00	97.66
FDR	6.31	5.40	4.65	3.88	2.66
F1-Score	93.72	94.62	95.49	96.05	97.50
MCC	87.47	89.25	90.99	92.12	95.00

Table 6.

Experimental analysis of the implemented model with new datasets using diverse classifiers.

Terms	Autoencoder³⁵	DTCN²⁹	BI-LSTM³⁰	DTCN-BI-LSTM³¹	ACP-CSA-DTCN-BI-LSTM
Dataset 2
Accuracy	93.51	93.70	94.86	95.55	97.15
Sensitivity	93.84	93.82	94.69	95.46	97.10
Specificity	93.17	93.58	95.03	95.64	97.21
Precision	93.26	93.64	95.05	95.66	97.22
FPR	6.83	6.42	4.97	4.36	2.79
FNR	6.16	6.18	5.31	4.54	2.90
NPV	93.75	93.77	94.67	95.44	97.08
FDR	6.74	6.36	4.95	4.34	2.78
F1-Score	93.55	93.73	94.87	95.56	97.16
MCC	87.01	87.40	89.72	91.10	94.31
Dataset 3
Accuracy	94.16	94.73	95.96	96.46	97.50
Sensitivity	94.21	94.87	95.89	96.61	97.66
Specificity	94.11	94.58	96.02	96.30	97.34
Precision	94.10	94.58	96.00	96.30	97.34
FPR	5.89	5.42	3.98	3.70	2.66
FNR	5.79	5.13	4.11	3.39	2.34
NPV	94.22	94.88	95.91	96.61	97.66
FDR	5.90	5.42	4.00	3.70	2.66
F1-Score	94.15	94.72	95.95	96.46	97.50
MCC	88.32	89.45	91.92	92.92	95.00

7. Conclusion

The developed multimodal fake news detection model based on deep learning was utilized to detect fake news publishers in society to support the public. The text data were gathered from the Twitter dataset. Those text data were taken to the pre-processing text process. The pre-process was done by stop words-punctuation removal and stemming. Then, those data were provided with the pre-processed text data as the output. Next, the pre-processed text data was carried out to fuse features. In that stage, image data was offered as the input. The VGG-16 model was utilized to extract the image features, and the TCNN model was supported to extract the text features. The features were selected optimally from those two models by the developed ACP-CSA. Moreover, the weights were optimized by the developed ACP-CSA. After that process, that provided the weighted fused features as the outcome. Then, the weighted fused features were taken as the input in the fake news detection process. There, parameters were tuned with the proposed DTCN-Bi-LSTM model by utilizing the developed ACP-CSA. Finally, the implemented fake news detection process provided the classified outcome. The best value of the implemented fake news detection process was 1.5% more than DO-DTCN-Bi-LSTM, 0.79% more than EOO-DTCN-Bi-LSTM, 1.5% improved than RSA-DTCN-Bi-LSTM and 0.29% enhanced than CSA-DTCN-Bi-LSTM. At last, the developed fake news model was performed more to provide the best values in all performances than other fake news detection models. Yet, the developed model suffers from a few complications where it needs to be resolved in the upcoming works. The analysis based on real-time data needs to be considered which is suggested to cope up with the developed model. The validation in real-time data ensures to offer accurate performance in the fake news detection process. In future work, the evaluation based on real-time data using the developed model will be implemented to detect fake news detection framework. Also, the ensemble model will be implemented.

Footnotes

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

References

Wei

Sun

, et al. Modality and event adversarial networks for multi-modal fake news detection. IEEE Signal Process Lett 2022; 29: 1382–1386.

Gupta

, et al. Combating fake news: stakeholder interventions and potential solutions. IEEE Access 2022; 10: 78268–78289.

Thota

Tilak

Ahluwalia

, et al. Fake news detection: a deep learning approach. SMU Data Sci Rev 2018; 1.

Umer

Imtiaz

Ullah

, et al. Fake news stance detection using deep learning architecture (CNN-LSTM). IEEE Access 2020; 8: 156695–156706.

de Oliveira

Medeiros

DSV

Mattos

DMF

. A sensitive stylistic approach to identify fake news on social networking. IEEE Signal Process Lett 2020; 27: 1250–1254.

Shrivastava

Kumar

Ojha

, et al. Defensive modeling of fake news through online social networks. IEEE Trans Comput Soc Syst 2020; 7: 1159–1167.

Sansonetti

Gasparetti

D’aniello

, et al. Unreliable users detection in social Media: deep learning techniques for automatic detection. IEEE Access 2020; 8: 213154–213167.

Nadeem

Ahmed

Zheng

, et al. SSM: Stylometric and semantic similarity oriented multimodal fake news detection. J King Saud Univ – Comput Inf Sci 2023; 35: 101559.

Shim

J-S

Lee

Ahn

. A link2vec-based fake news detection model using web search results. Expert Syst Appl 2021; 184: 115491.

10.

Silva

Han

Luo

, et al. Propagation2Vec: Embedding partial propagation networks for explainable fake news early detection. Inf Process Manag 2021; 58: 102618.

11.

Samadi

Mousavian

Momtazi

. Deep contextualized text representation and learning for fake news detection. Inf Process Manag 2021; 58: 102723.

12.

Paka

Bansal

Kaushik

, et al. Cross-SEAN: A cross-stitch semi-supervised neural attention model for COVID-19 fake news detection. Appl Soft Comput J 2021; 107: 107393.

13.

Nasir

Khanb

Varlamis

. Fake news detection: A hybrid CNN-RNN based deep learning approach. Int J Inf Manag Data Insights 2021; 1: 100007.

14.

Trueman

, et al. Attention-based C-BiLSTM for fake news detection. Appl Soft Comput 2021; 110: 107600.

15.

Kumari

Ekbal

. AMFB: Attention based multimodal Factorized Bilinear Pooling for multimodal Fake News Detection. Expert Syst Appl 2021; 184: 115412.

16.

Meel

Vishwakarma

. HAN, image captioning, and forensics ensemble multimodal fake news detection. Inf Sci 2021; 567: 23–41.

17.

Sheikhi

. An effective fake news detection method using WOA-xgbTree algorithm and content-based features. Appl Soft Comput 2021; 109: 107559.

18.

Jiang

Haq

, et al. A novel stacking approach for accurate detection of fake news. IEEE Access 2021; 9: 22626–22639.

19.

Kao

H-Y

. MVAN: multi-view attention networks for fake news detection on social Media. IEEE Access 2021; 9: 106907–106917.

20.

Dong

Victor

Qian

. Two-path deep semisupervised learning for timely fake news detection. IEEE Trans Comput Social Syst 2020; 7: 1386–1398.

21.

Song

Shu

. Temporally evolving graph neural network for fake news detection. Inf Process Manag 2021; 58: 102712.

22.

Jing

Sun

, et al. Multimodal fake news detection via progressive fusion networks. Inf Process Manag 2023; 60: 103120.

23.

Song

Ning

Zhang

, et al. A multimodal fake news detection model based on crossmodal attention residual and multichannel convolutional neural networks. Inf Process Manag 2021; 58: 102437.

24.

Kaliyar

Goswami

Narang

. Echofaked: improving fake news detection in social media with an efficient deep neural network. Neural Comput Appl 2021; 33: 8597–8613.

25.

Agarwal

Mittal

Pathak

, et al. Fake news detection using a blend of neural networks: an application of deep learning. SN Comput Sci 2020; 143.

26.

Braik

. Chameleon swarm algorithm: A bio-inspired optimizer for solving engineering design problems. Expert Syst Appl 2021; 174: 114685.

27.

Tammina

. Transfer learning using VGG-16 with deep convolutional neural network for classifying images. Int J Sci Res Publ 2019; 9: 143–150.

28.

Soni

Chouhan

Rathore

. Textconvonet: a convolutional neural network based architecture for text classification. Appl Intell 2023; 53: 14249–14268.

29.

Shi

Lin

, et al. Deep attention gated dilated temporal convolutional networks with intra-parallel convolutional modules for End-to-End monaural speech separation. Interspeech 2019: 15–19.

30.

Shao

, et al. Domain-specific Chinese word segmentation based on bi-directional long-short term memory model. IEEE Access 2019; 7: 12993–13002.

31.

Jiang

. Prediction of air pollutant concentrations based on TCN-BiLSTM-DMAttention with STL decomposition. Sci Rep 2023; 13.

32.

Almazán-Covarrubias

Peraza-Vázquez

Peña-Delgado

, et al. An improved dingo optimization algorithm applied to SHE-PWM modulation strategy. Appl Sci 2022; 3.

33.

Azaki

Cresswell

. Level of local human disturbance and feeding state determines escape behaviour in eurasian oystercatchers. Ethology 2021; 3127: 986–994.

34.

Elgamal

Sabri

AQM

Tubishat

, et al. Improved Reptile search optimization algorithm using chaotic map and simulated annealing for feature selection in medical field. IEEE Access 2022; 10: 51428–51446.

35.

Park

Cho

D-H

. ConvAE: a new channel autoencoder based on convolutional layers and residual connections. IEEE Commun Lett 2019; 23: 1769–1772.