A New Model for a Secure Social Media Application

Abstract

The progress of data technology and wireless networks is generated by open online communication channels. Unfortunately, trolls are abusing the technology for executing cyberattacks and threats. An automated cybersecurity solution is vital for avoiding the threats and security issues from social media. This can be a requirement for tackling and considering cyberbullying in various aspects including prevention of such incidents and automated detection. This study introduces a novel Artificial Fish Swarm Algorithm with Weighted Extreme Learning Machine (AFSA-WELM) model for cybersecurity on social media. The proposed model is mostly intended to detect the existence of cyberbullying on social media. The proposed model starts by processing the dataset and making it ready for the next stages of the model. It then uses the TF-IDF vectorizer for word embedding. After that, it uses the WELM model for the identification and classification of cyberbullying. Finally, the optimal tunning parameters used in the WELM model are derived for the AFSA model. The experimental analysis has shown that the proposed model achieves maximum accuracy compared with existing algorithms. Moreover, our proposed model achieves maximum precision–recall performance with various datasets.

1. Introduction

Cyberbullying has been a progressive and always developing issue all through most recent years, and thus programmed recognition of it has been an engaging point for scientists. Cyberbullying or disdain discourse is the demonstration of posting scornful substance on social media, which is found to be harmful because of its tedious and oppressive nature [1]. Cyberbullying can influence individuals’ mental states and different parts of life. Particularly for youngsters, cyberbullying could even lead them to cause self-harm and self-destruction [2]. Scientists have likewise concentrated on the relationship between cyberbullying exploitation and the risk of self-destructive ideation [3,4]. It is a fundamental and provoking task to distinguish cyberbullying. Concerning its significance, although the mischief of cyberbullying should be limited as much as can be expected, it is more vital to foster a programmed model to foresee cyberbullying occasions, rather than checking physically or helping victims a while later.

The difficulties of cyberbullying discovery are as follows [5]. To begin with, according to the point of view of manual verification, the judgment of whether specific conduct is cyberbullying changes from one individual to another. For instance, when a sentence contains swear words, it may very well be considered cyberbullying activity, yet now and again, this is not accurate, such as when some teenagers speak with their companions on social networks. After that, where cyberbullying happens most often [6], social organization’s platforms generally have the idea of communicating with the public and unknown commenters. Thus, posts are more autonomous and context-free, and it is simple to cause doubt or misdirect others. Third, one significant issue with cyberbullying research is the absence of standard information [7]. Although the information utilized in numerous past investigations was obtained from similar social networks (e.g., Twitter), they were made by freely utilizing public API or by site scraping. Subsequently, individual data cannot be measured against each another and are not helpful for checking the technique’s comprehensiveness.

This study introduces a novel Artificial Fish Swarm Algorithm with Weighted Extreme Learning Machine (AFSA-WELM) model for cybersecurity on social media. The AFSA-WELM model presented here is mainly intended to detect the existence of cyberbullying on social media. The main contributions of this study are summarized as follows:

• Introducing an efficient AFSA-WELM model to detect cyberbullying on social media using an artificial fish swarm algorithm and weighted extreme machine learning.
• Using the TF-IDF model for feature extraction, using the WELM model for classification, and applying the AFSA model for parameter tuning.
• The presented AFSA-WELM model has been validated through extensive experimental evaluations on various datasets. It achieved the maximum precision–recall performance for various datasets.

Finally, the presented AFSA-WELM has shown maximum accuracy compared with existing algorithms such as the SVM and Naïve Bayes algorithms.

The rest of this study is organized as follows. Section 2 presents the related work and discussion. Section 3 introduces the proposed model. Section 4 presents the detailed results and discussion. This is followed by the conclusions in Section 5.

2. Related Work

Regarding instant messages in social networks, the length of sentences is generally short and noisy. These messages can likewise be unstructured, e.g., messages could contain emoticons, emojis, or incorrect spellings, which befuddles models catching information from instant messages [8]. While a multitude of techniques have attempted to work on the exactness of the cyberbullying-related task somewhat, we accept that they have not completely used precise data on past text-based information, which we call the setting in the context of this study [9]. The main classification of setting is an inner setting that incorporates any non-text-based data that can be separated from the dataset itself, e.g., the client’s liked posts, adherents, or multimedia content connected to the social media post, e.g., pictures and recordings [10,11].

Bozyiğit et al. [12] described the significance of social networking characteristics in detecting cyberbullying. Initially, a balanced dataset consisting of 5000 labeled contents with numerous social networking characteristics was built. Next, the relationships between social media features and cyberbullying were analyzed by chi-square testing. It is obvious that a few characteristics (for example, the sender’s followers) are powerfully connected with online bullying events according to the tested outcomes. The researchers in [13] proposed a DNN method for detecting cyberbullying in three distinct types of social information, such as visual media, infographics, and textual content (embedded text and an image). The compact structure, CapsNet–ConvNet, comprises ConvNet to predict the visual bullying content and CapsNet-DNN with dynamic routing to predict text bullying content. The researchers in [14] presented an ML technique that (i) fine-tunes variants of BERT, a deep attention-based language technique that can detect patterns in a long and noisy body of text; (ii) extracts contextual data from various sources involving external knowledge sources, metadata, and images, and uses those characteristics for complementing the learning method; and (iii) effectively integrates textual and contextual characteristics through boosting and a wide-ranging framework. Fang et al. [15] presented the advantages of GRU cells and Bi-GRU for learning the fundamental relationships among words in these two directions. Rupesh et al. [16] used machine learning algorithms to efficiently detect hate speech on social networks. In addition, we presented a self-attention model which had the advantage of achieving great efficiency in cyberbullying classification tasks.

In this study, we conducted a comparative analysis of the presented the proposed model with some recent models, as shown in

Table 1. A comparative analysis of the AFSA-WELM technique and existing algorithms.

Methods	Accuracy	Precision	Recall
SVM (linear algorithm)	95.17	93.75	95.07
SVM (poly-algorithm)	98.55	97.94	98.79
SVM (RBF algorithm)	87.28	88.57	89.67
SVM (sigmoid algorithm)	98.29	98.43	98.50
Naïve Bayes algorithm	95.59	95.24	96.70
The proposed model	99.14	95.22	99.53

.

The table portrays a brief comparative study of the AFSA-WELM model and recent models [17].

3. The Proposed Model

This study introduces a novel model for cybersecurity on social media, as shown in Figure 1. The figure illustrates the overall process of the proposed model. The presented model is mainly intended to detect the existence of cyberbullying on social media. The proposed model is composed of five main stages:

• Data input: The model accepts the training dataset.
• Data pre-processing: The inputted data will be pre-processed to the formats that will be used in the next stages.
• Feature extraction: The model uses a TF-IDF vectorizer for word embedding.
• Classification and identification: The model uses WELM to efficiently identify and classify cyberbullying.
• Parameter tunning: In this stage, the AFSA model is derived to perfectly obtain the optimal tuned parameters that are used in the proposed model.

Finally, in order to validate the introduced model, we introduced a comparative study that showed the efficiency of the introduced model over other methods in the related literature.

3.1. TF-IDF Vectorizer

In this work, the TF-IDF vectorizer was used for word embedding. TF-IDF is an extensively used feature extraction system for examining texts. The two substantial tasks are indexing and weighting; TF-IDF manages the weighting. It determines the weight of t in a document D. TF-IDF is derived from TF and IDF as follows [18]:

$$TF\left(t\right)=\frac{t_D}{N_D}$$

(1)

$$IDF\left(t\right)=log\frac{d}{dt}$$

(2)

Now $t_D,$ $d$, and $dt$ are used to characterize the inclusive number of times of $t$ appears in a document $D$, the total number of documents and the quantity of documents with term $t.$

The weight of each term according to the TF-IDF is estimated by:

$$W_{t,d}=T{F}_{t,d}\left(\frac{t_D}{d_{f,t}}\right)$$

(3)

where $T{F}_{t,d}$ and $d_{f,t}$ denote the frequency of $t$ in document $d$ and the sum of documents containing $t$, respectively.

3.2. WELM-Based Classification

Next, the WELM model was used for the identification and classification of cyberbullying. The WELM classification model was used to assign appropriate class labels. ELM was used for the classification of a balanced dataset, while WELM was used for the classification of an imbalanced dataset [19]. The training dataset has $N$ distinct instances$, \left(x_i, z_i\right),i=1,2,\dots ,N$. A single hidden layer NN with $L$ hidden layer node was formulated by:

$${\displaystyle \sum }_{i=1}^L{\beta }_i\cdot 1\left(w_i\cdot x_j+b_i\right)=z_j,j=1,\dots ,N$$

(4)

whereas $w_i$ characterizes a single hidden layer’s input weight, $l()$ denotes the activation functions, ${\beta }_i$ indicates the output weights, and $b_i$ signifies the bias of the individual hidden layers.

$$S\beta =T$$

(5)

where $S$ denotes the output matrix of a single hidden layer

$$S\left(w_1,\dots ,w_L,b_1,\dots ,b_L,x_1,\dots ,x_N\right)={\left(\begin{array}{ccc}\mathrm{l}(w_1\cdot x_1+b_1& \cdots & \mathrm{l}\left(w_L{x}_1+b_L\right)\\ ⋮& \ddots & ⋮\\ \mathrm{l}\left(w_1\cdot x_N+b_1\right)& \cdots & \mathrm{l}\left(w_L\cdot x_N+b_L\right)\end{array}\right)}_{N\times L}$$

(6)

Based on the Karush–Kuhn–Tucker concept, the Lagrangian factor was used to change the trained ELM as a dual problem [20]. The output weight $\beta$ can be evaluated as follows:

$$\beta =S^T{\left(\frac{1}{C}+S{S}^T\right)}^{-1}T$$

(7)

Here, $C$ denotes the regularization coefficient. Therefore, the output function of the ELM classification is formulated by:

$$F\left(x\right)=s\left(x\right)S^T{(\frac{1}{C}+S{S}^T)}^{-1}T={\left[\begin{array}{c}K\left(x,x_1\right)\\ ⋮\\ K\left(x,x_N\right)\end{array}\right]}^T{\left(\frac{1}{C}+\chi \right)}^{-1}T$$

(8)

where $\chi$ denotes the kernel matrix, which is calculated by:

$$\chi =S{S}^T=s\left(x_i\right)s\left(x_j\right)=K\left(x_i,x_j\right)$$

(9)

It is obvious in (13) that the hidden layer feature maps $s\left(x\right)$ denote the independence in classification results of ELM. The classification result is related to the kernel function, $K\left(x,y\right)$. $K\left(x,y\right)$ denotes an inner product, and the number of hidden layers is no outcomes in the resulting output.

$$K\left(x,y\right)=\exp \left(-\gamma \Vert x-y{\Vert }^2\right)$$

(10)

Therefore, the KELM classification efficacy is distributed into two variables, namely the kernel function parameter, $\gamma$ and the penalty parameter, C. The original goal of ELM is achieved in WELM by preserving the weight for distinct instances to handle imbalanced classification problems. It is calculated by the following equations:

$$F\left(x\right)={\left[\begin{array}{c}K\left(x,x_1\right)\\ ⋮\\ K\left(x,x_N\right)\end{array}\right]}^T{\left(\frac{1}{C}+W\chi \right)}^{-1}WT$$

(11)

$$W=diag \left(w_{ii}\right),i=1,2, \dots , N$$

(12)

where $W$ denotes the weight matrix. WELM has two weightage models as follows:

$$w_{ii}=\frac{1}{\#\left(z_i\right)}$$

(13)

$$w_{ii}=\{\begin{array}{l}\frac{0.618}{\#\left(z_i\right)},if z_i>\overline{z}\hfill \\ \frac{1}{\#\left(z_i\right)},otherwise\hfill \end{array}$$

(14)

where $\#\left(z_i\right)$ denotes the number of instances assigned to class $z_i,$ $i=1,\dots ,m$; $m$ denotes the number of classes; and $\overline{z}$ indicates the average instance of all the classes. Figure 2 showcases the framework of ELM.

3.3. AFSA-Based Parameter Optimization

At this stage, the AFSA model is derived for optimal tuning of the parameters involved in the WELM model. The ASFA replicates the behavior of a prey fish swarm and it has been established for rapid searching. Exploration begins with arbitrary result maps for each fish. Fitness is assessed and the fish following techniques are initiated, and each fish follows the other fishes via the optimum outcome. Once the outcome is sub-optimum, a swarming method is initiated. Once the chosen outcome is not linked, the prey method is initiated. This process repeats until the required threshold or termination condition is encountered [21]. The AFSA has four functions that have been modeled on the nature of fish in a fish swarm. A primary function is free-moving action: in the environment, a fish travels in a swarm when it is not injured. The next function is prey action, since it uses smell, vision, and other available senses in the body. In AFSA, the area where an artificial fish senses a predator is modeled as a neighborhood via a visual-sized radius. The next function is following action when a fish identifies food; other swarm members follow it to reach the food. The last function is the swarm action, which replicates fishes as each fish continually attempts to be in the swarm and does not leave so it can be protected from the hunter.

Consider the condition vector of AFS is $\left(x_1,x_2,\dots x_n\right)$, where $x_1,x_2,\dots x_n$ denotes the location of the fish. The food in this location of the fish is characterized by $y=f\left(x\right)$, in which $y$ indicates the objective function value. The distance amongst artificial fish is $d_{ij}=\Vert X_i-X_j\Vert ,\mathrm{where}i$ and $j$ represent arbitrary fish, Step indicates the maximum step size of artificial fish, and $a$ denotes the degree of congestion factor.

Let $X_v$ be the visual location at a particular moment. $X_{next}$ denotes the novel location. The development process is shown as follows:

$$X_v=X_i+Visual\times rand(),i\in \left[0,n\right]$$

(15)

$$X_{next}=X+\frac{X_v-X}{\Vert X_v-X\Vert }\times step\times rand()$$

(16)

During the prey model, the behavior of fish is represented as follows:

$$prey \left(X_i\right)=\{\begin{array}{cc}{x}_i+step\frac{x_j-x_i}{\Vert x_j-x_i\Vert }\hfill & if y_j-y_i\hfill \\ x_i+\left(2rand-1\right).step\hfill & else\hfill \end{array}$$

(17)

where rand denotes a random function within $\left[0,1\right].$In the swarming phase, the behavior of the swarm is represented as follows:

$$swarm\left(X_i\right)=\{\begin{array}{cc}{X}_i+step\frac{x_j-x_i}{\Vert x_j-x_i\Vert }\hfill & if\frac{y_c}{nf}>\delta y_i\hfill \\ prey\left(X_i\right)\hfill & else\hfill \end{array}$$

(18)

The next stage is represented in the following (Equation (19)):

$$follow \left(X_i\right)=\{\begin{array}{cc}{X}_i+step\frac{x_{max}-x_i}{\Vert x_{max}-x_i\Vert }\hfill & if\frac{y\max }{nf}>\delta y_i\hfill \\ prey\left(X_i\right)\hfill & else\hfill \end{array}$$

(19)

where x_max is the last location that the fish can arrive and y_max is the maximum value of the objective function to guarantee that local and global exploration, and the exploration direction towards the optimum food source are carried out.

4. Results and Discussion

The presented AFSA-WELM model was simulated using a dataset collected from the Kaggle repository. The dataset holds 12,729 samples with 11,661 data classified as non-cyberbullying and 1068 data in the cyberbullying class.

Figure 3 illustrates the confusion matrices offered by the AFSA-WELM model for distinct sizes of training/testing datasets. With a training/testing data split of 90:10, the proposed model identified 1158 samples in the non-cyberbullying class and 104 samples in the cyberbullying class. In addition, with a training/testing data split of 80:20, the AFSA-WELM method identified 2323 samples in the non-cyberbullying class and 187 samples in the cyberbullying class. Meanwhile, with a training/testing data split of 70:30, the proposed model identified 3432 samples in the non-cyberbullying class and 215 samples in the cyberbullying class. Finally, with a training/testing data split of 60:40, the AFSA-WELM technique identified 4619 samples in the non-cyberbullying class and 347 samples in the cyberbullying class.

Table 2. Results of the proposed model with different measures and training/testing datasets sizes.

Class Labels	Accuracy	Precision	Recall	F-Score
Training/Testing (90:10)
Non-Cyberbullying	99.14	100.00	99.06	99.53
Cyberbullying	99.14	90.43	100.00	94.98
Average	99.14	95.22	99.53	97.25
Training/Testing (80:20)
Non-Cyberbullying	98.59	99.02	99.44	99.23
Cyberbullying	98.59	93.50	89.05	91.22
Average	98.59	96.26	94.25	95.23
Training/Testing (70:30)
Non-Cyberbullying	95.50	96.98	98.14	97.56
Cyberbullying	95.50	76.79	66.77	71.43
Average	95.50	86.88	82.46	84.49
Training/Testing (60:40)
Non-Cyberbullying	97.53	98.32	98.99	98.65
Cyberbullying	97.53	88.07	81.46	84.63
Average	97.53	93.19	90.22	91.64

reports the overall classifier results of our proposed model with different training/testing dataset sizes. With a training/testing dataset split of 90:10, the AFSA-WELM model resulted in an $acc{u}_y$, $pre{c}_n$, $rec{a}_l$, and $F_{score}$ of 99.14%, 95.22%, 99.53%, and 97.25%, respectively. Moreover, with a training/testing dataset split of 80:20, the proposed model resulted in an $acc{u}_y$, $pre{c}_n$, $rec{a}_l$, and $F_{score}$ of 98.59%, 96.26%, 94.25%, and 95.23%c correspondingly. In addition, with a training/testing dataset split of 70:30, the proposed model resulted in an $acc{u}_y$, $pre{c}_n$, $rec{a}_l$, and $F_{score}$ of 95.50%, 86.88%, 82.46%, and 84.49%, correspondingly. Lastly, with a training/testing dataset split of 60:40, the proposed model resulted in an $acc{u}_y$, $pre{c}_n$, $rec{a}_l$, and $F_{score}$ of 97.53%, 93.19%, 90.22%, and 91.64%, correspondingly.

A brief precision–recall examination of the AFSA-WELM model with different training/testing dataset sizes is portrayed in Figure 4. By observing the figure, it can be noticed that the AFSA-WELM model achieved maximum precision–recall performance for all datasets. The figure shows the recall and the precision values of the four different samples.

Figure 5 portrays the ROC investigation of the proposed model with different training/testing dataset sizes. The figure shows the false positive rate and the true positive rate. The figure has four different training/testing samples to verify the stability of the proposed model’s performance. The figure shows that the AFSA-WELM model produces adequate results with maximum ROC values for the different class labels.

Figure 6 reports a comparative $acc{u}_y$ investigation of the AFSA-WELM model and existing models. The figure shows that the SVM (RBF algorithm) achieved a lower $acc{u}_y$ of 87.28%. Next, the SVM (linear algorithm) and the Naïve Bayes algorithm achieved a slightly increased $acc{u}_y$ of 95.17% and 95.59%, respectively. Along with that, the SVM (poly-algorithm) and SVM (sigmoid algorithm) models achieved a reasonable $acc{u}_y$ of 98.55% and 98.29%, respectively. However, the AFSA-WELM model achieved the maximum $acc{u}_y$ of 99.14%.

Figure 7 shows a comparative $pre{c}_n$ investigation of our proposed model and existing models. The figure portrays that the SVM (RBF algorithm) obtained a lower $pre{c}_n$ of 88.57%. Moreover, the SVM (linear algorithm) and the Naïve Bayes algorithm attained a slightly increased $acc{u}_y$ of 93.75% and 95.24%, respectively. Next, the SVM (poly-algorithm) and SVM (sigmoid algorithm) models reached a reasonable $pre{c}_n$ of 97.94% and 98.43%, correspondingly. Finally, the proposed method achieve a maximal $acc{u}_y$ of 95.22%.

Figure 8 showcases a comparative $rec{a}_l$ investigation of the proposed method and existing models. The figure shows that the SVM (RBF algorithm) achieved a lower $rec{a}_l$ of 89.67%. Moreover, the SVM (linear algorithm) and the Naïve Bayes algorithm reached a somewhat increased $rec{a}_l$ of 95.07% and 96.70%, correspondingly. At the same time, the SVM (poly-algorithm) and SVM (sigmoid algorithm) models reached a reasonable $rec{a}_l$ of 98.795% and 98.43%, correspondingly. Finally, the proposed model achieved a maximum $rec{a}_l$ of 99.53%.

After examining the results and discussion, it was verified that the proposed model resulted in maximum performance in terms of the detection and classification of cyberbullying.

5. Conclusions

This study presented a novel model for cybersecurity on social media. The proposed model developed here mostly detects the existence of cyberbullying on social media. The introduced model is composed of five main stages: input, pre-processing, feature extraction, classification and identification, and parameter tuning. Our model uses an artificial fish swarm algorithm and weighted extreme machine learning to effectively detect cyberbullying on social networks. It also uses the TF-IDF model to successfully carry out the feature extraction process with the dataset. Moreover, it uses WELM and AFSA for classification and parameter tuning, respectively. The experimental results show that the developed method achieved a maximum $acc{u}_y$ and $rec{a}_l$ in comparison with the SVM and Naïve Bayes algorithms. Future extensions of this study are as follows. Deep learning models can be introduced to improve the cyberbullying detection outcomes. Moreover, an extensive comparison needs to be conducted between the proposed model and some recent models in the related literature.