(IJACSA) International Journal of Advanced Computer Science and Applications,

Vol. 14, No. 7, 2023

Ensemble Deep Learning (EDL) for Cyber-bullying

on Social Media
Zarapala Sunitha Bai1*, Sreelatha Malempati2
Department of Computer Science and Engineering-Y.S.R University College of Engineering & Technology,
Acharya Nagarjuna University, Guntur 522510, Andhra Pradesh, India1
Department of Computer Science and Engineering, R.V.R & J.C College of Engineering,
Chowdavaram, Guntur-522019, India2

Abstract—Cyber-bullying is a growing problem in the digital identify the sentiment expressed in a text, whether it is
age, affecting millions of people worldwide. Deep learning positive, negative, or neutral [2]. One approach to using
algorithms have the potential to assist in identifying and sentiment analysis for cyber-bullying detection is to look for
combating Cyber-bullying by detecting and classifying harmful negative sentiments expressed towards an individual or group,
messages. This paper uses two Ensemble deep learning (EDL) such as derogatory or offensive language, insults, or threats
models to detect Cyber-bullying on text data, images and [3]. These can be identified using NLP techniques, such as
videos—and an overview of Cyber-bullying and its harmful part-of-speech tagging, named entity recognition, and
effects on individuals and society. The advantages of using deep sentiment analysis algorithms. Another approach is to analyze
learning algorithms in the fight against Cyber-bullying include
the context in which the message is being conveyed, such as
their ability to process large amounts of data and learn and
adapt to new patterns of Cyber-bullying behaviour. For text
the topic being discussed and the online community it is being
data, firstly, a pre-trained model BERT (Bidirectional Encoder shared in [4] [5]. For example, if a message contains negative
Representations from Transformers) is used to train on cyber- sentiments towards a specific group of people or individual,
bullying text data. The next step describes the data pre- and is being shared in an online community known for hostile
processing and feature extraction techniques required to prepare or aggressive behaviour, it may be a sign of cyber-bullying. It
data for deep learning algorithms. We also discuss the different is important to note that sentiment analysis is not a foolproof
types of deep learning algorithms that can be used for Cyber- method for detecting cyber-bullying and should be used in
bullying detection, including convolutional neural networks combination with other techniques, such as human moderation
(CNNs), recurrent neural networks (RNNs), and deep belief and reporting mechanisms. Additionally, it is important to
networks (DBNs). This paper combines the sentiment analysis ensure that the use of sentiment analysis does not infringe on
model, such as Aspect-based Sentiment Analysis (ABSA), for individuals' privacy rights or result in false accusations.
classifying bullying messages. Deep Neural network (DNN) used
the classification of Cyber-bullying images and videos. Cyber-bullying (CB) is a default model for publicly
Experiments were conducted on three datasets such as Twitter abusing a person. Many online social media networking
(Kaggle), Images (Online), and Videos (Online). Datasets are (OSMN) like Facebook, Twitter, and Instagram act as a
collected from various online sources. The results demonstrate medium for people based on cyber-bullying attacks [6].
the effectiveness of EDL and DNN in detecting Cyber-bullying in Several automated models aim to develop to classify cyber-
terms of detecting bullying data from relevant datasets. The EDL bullying in terms of text messages, audio, and videos [7].
and DNN obtained an accuracy of 0.987, precision of 0.976, F1- Sometimes based on the topic modeling, cyber-bullying
score of 0.975, and recall of 0.971 for the Twitter dataset. The attacks occur in several datasets belonging to topic modeling.
performance of Ensemble CNN brought an accuracy of 0.887, Twitter has become more popular for cyber-bullying by using
precision of 0.88, F1-score of 0.88, and recall of 0.887 for the various types of attacks. Exemplary-grained automated
Image dataset. For the video dataset, the performance of models were developed to detect cyber-bullying regarding
Ensemble CNN is an accuracy of 0.807, precision of 0.81, F1- topic modeling [8]. It is essential to complain the cyber-
score of 0.82, and recall of 0.81. Future research should focus on bullying attacks in OSMS if the user violates the ITE Law,
developing more accurate and efficient deep learning algorithms which is considered a crime in OSMS like Twitter [9]. The
for Cyber-bullying detection and investigating the ethical
victim should act if any abusive language is used on Twitter.
implications of using such algorithms in practice.
The aim of cyber-bullying mainly focuses on classifying the
Keywords—Cyber bullying; ensemble deep learning (EDL); tweets present on Twitter. This paper describes the automated
convolutional neural networks (CNNs); recurrent neural networks approach for detecting cyber-bullying attacks by using the
(RNNs); deep belief networks (DBNs) sentiment analysis on text messages, videos, and audio.
Sentiment analysis helps the proposed automated approach to
I. INTRODUCTION find the cyber-bullying attacks in multi-media. An aspect
Sentiment analysis can be used to detect instances of based sentiment analysis model combined with automated
cyber-bullying by analyzing the language and tone used in classification approach used to classify the cyber-bullying
online messages, comments, or posts [1]. The process involves words from given input data. It shows the improved
using natural language processing (NLP) techniques to
*Corresponding Author.

551 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 14, No. 7, 2023

performance in terms of accuracy, sensitivity, specificity, F1- al. [12] developed a fine-grained model to detect cyber-
score and precision. bullying messages based on linguistic analysis. The proposed
model focused on finding the patterns based on Linguistic
The organization of the paper is as follows. Section II Inquiry Word Count (LIWC) to detect fine-tuned cyber-
explains the literature survey about various methods of bullying detection from different social media datasets. Dalvi
cyberbullying. Section III and Section IV give the significance et al. [13] proposed an ML model to see cyber-bullying from
of the work and training and feature extraction techniques. social media posts like Twitter. The proposed ML model is
Section V describes the experimental results with comparative used to prevent bullying on Twitter. Using the Twitter API,
performance. Section VI describes the conclusion. the tweets are extracted and classified whether the tweets are
A. Significant Points of Proposed System bullied or not. Zhao et al. [14] proposed a new learning model
Cyber-bullying has become a common problem in today's to solve the issues in cyber-bullying detection—the proposed
digital age, with severe psychological and emotional approach combined with a deep learning model to denoising
consequences for victims. Detecting and preventing cyber- auto-encoder (SDA). The SDA model contains semantic
bullying is critical for ensuring individuals' well-being in dropout noise and sparsity. These features focused on
online communities. Manual monitoring of online content, on knowledge and the word embedding method. The performance
the other hand, is a time-consuming and inefficient process. of the proposed model improved by combining it with
As a result, there is a need to create an automated system that semantic-enhanced marginalized (SEM) to find the hidden
uses deep learning models to detect cyber-bullying. This features of the bullying content. The version of the proposed
project aims to develop and test a deep-learning model to model was analyzed using two datasets such as Twitter and
detect cyber-bullying in online text, images, and videos. The MySpace, and achieved better performance in terms of
model should be able to classify messages, comments, or posts classification. Luo et al. [15] proposed the BiGRU-CNN for
as cyber-bullying or non-cyber-bullying based on their classifying cyber-bullying messages. BiGRU mainly focused
content. The detection system's accuracy, precision, and recall on extracting the global features that significantly impact
should be high, with low error rates. organizing bullying messages. The CNN consists of a
convolution method with 128 kernels of length 5; this is used
II. LITERATURE SURVEY to extract the features that improve the learning rate of the
model better. Adav et al. [16] introduced the BERT model,
Semantic-enhanced marginalized denoising auto-encoder which is suitable for creating contextual embeddings that
(SEDMA) is a type of neural network that is trained to produce the particular embeddings for classifying cyber-
reconstruct clean data from noisy data by learning the bullying detection in social media. Ahmed et al. [17]
underlying distribution of the input data. It has been enhanced introduced the cyber-bullying model that classifies the
with semantic information to improve its ability to capture the bullying words belonging to Bangla and Romanized Bangla
context and meaning of the text [10]. To use SEDMA for texts utilizing ML and DL models. Iwendi et al. [18]
cyber-bullying detection, the first step is to train the model on performed various DL models that detect cyber-bullying in
a dataset of labelled examples of cyber-bullying and non- social media. The comparison between different DL
cyber-bullying text. During training, the SEDMA learns to algorithms shows high performance. Aind et al. [19]
identify patterns and features that distinguish between cyber- introduced the novel Q-Bully model that sees cyber-bullying
bullying and non-cyber-bullying text. Once the SEDMA is automatically from social media platforms. The proposed
trained, it can be used to detect cyber-bullying in new text. performance is improved by combining with Reinforcement
The input text is first pre-processed to remove noise and Learning gives better accuracy. Ketsbaia et al. [20] introduced
convert it into a numerical representation that can be input to the DL models that detect cyber-bullying automatically.
the SEDMA. The SEDMA then reconstructs the clean version Pradhan et al. [21] proposed the new DL model that sees the
of the input text, and the reconstruction error is used to cyber-bullying from Wikipedia, Formspring, and Twitter
determine whether the input text is cyber-bullying or not. The cyber-bullying datasets.
advantage of using SEDMA for cyber-bullying detection is
that it can capture the semantic meaning of the text, which is III. HOW CYBER-BULLYING AFFECTS THE SOCIAL MEDIA
critical for detecting subtle forms of cyber-bullying.
Additionally, it is more robust to noise and can handle Cyber-bullying can have a significant impact on social
variations in the input text. In conclusion, the use of semantic- media use, both for individuals and as a whole. Here are some
enhanced marginalized denoising auto-encoder is a promising of the ways it can affect social media introduction:
approach for cyber-bullying detection, and it has the potential 1) Fear of harassment: Cyber-bullying can create a
to improve the accuracy of current cyber-bullying detection climate of fear on social media platforms. Individuals who
systems. Zhang et al. [11] proposed the novel pronunciation- have been bullied in the past or who fear being bullied may be
based CNN to solve issues in detecting cyber-bullying based
hesitant to join social media or may limit their use of these
on the pronunciation of misspelled words. The proposed
approach corrects the errors that occur by spellings that didn't platforms.
change in accent because of its noise and data sparsity 2) Damage to reputation: Cyber-bullying can damage an
imbalance present in the dataset. To solve these issues, the individual's reputation, making them less likely to want to be
proposed model applied to two datasets, Twitter and Form active on social media. This can also affect the reputation of
spring. The proposed approach's comparative performance the platform itself if it becomes known as a place where
shows more effective results than existing models. Zhang et bullying is rampant.

552 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 14, No. 7, 2023

3) Decreased engagement: Cyber-bullying can lead to 4) Reduced trust: If social media platforms are seen as a
decreased engagement on social media, as individuals may place where cyber-bullying is common, users may lose trust in
avoid posting or interacting with others out of fear of being these platforms and be less likely to use them. This can affect
targeted. This can have a negative impact on the platform as a the growth and sustainability of social media as a whole.
whole, as it relies on user engagement to generate revenue. Overall, cyber-bullying can have a significant impact on social
4) Reduced trust: If social media platforms are seen as a media use and adoption, and it is important for individuals and
place where cyber-bullying is common, users may lose trust in social media companies to take steps to prevent and address
these platforms and be less likely to use them. This can affect this issue.
the growth and sustainability of social media as a whole. 5) Here are the equations for the fine-tuning process:
Overall, cyber-bullying can have a significant impact on social First, we add a classification layer on top of the pre-trained
media use and adoption, and it is important for individuals and BERT model. The classification layer consists of a fully
social media companies to take steps to prevent and address connected layer and a soft-max activation function.
this issue, Fig. 1.
, - (1)
( ) (2)
Where is the hidden state of the [CLS] token,
and are the weight and bias parameters of the fully
connected layer, and is the predicted probability
distribution over the two classes (cyber-bullying and non-
cyber-bullying).
We then define the loss function as the cross-entropy loss
between the predicted and true labels:
∑ ( ) ( ) ( )) (3)
Where is the true label (1 for cyber-bullying, 0 for non-
cyber-bullying), and is the predicted probability for the
ith message.
We optimize the parameters of the classification layer by
minimizing the loss function using gradient descent:
Fig. 1. Types of cyber-bullying.
( ) (4)
IV. CYBER-BULLYING DETECTION MODEL FOR TEXT The above equations outline the fine-tuning process for
MESSAGES BERT in cyber-bullying detection. Note that this process
A. BERT (Bidirectional Encoder Representations from requires a labelled dataset of cyber-bullying and non-cyber-
Transformers) for Training on Cyber-Bullying Data bullying messages, as well as appropriate pre-processing and
tokenization of the input text.
Cyber-bullying can have a significant impact on social
media use, both for individuals and as a whole. Here are some B. Pre-processing Techniques for Cyber-Bullying
of the ways it can affect social media introduction: 1) Tokenization: Tokenization divides the text into smaller
1) Fear of harassment: Cyber-bullying can create a units known as tokens, which can be words, phrases, or
climate of fear on social media platforms. Individuals who symbols. It is a common step in pre-processing natural
have been bullied in the past or who fear being bullied may be language processing (NLP) tasks such as sentiment analysis
hesitant to join social media or may limit their use of these and topic modelling. Consider the following example sentence
platforms. to demonstrate tokenization for cyber-bullying data:
2) Damage to reputation: Cyber-bullying can damage an  "I hate you and wish you were never born. You're
individual's reputation, making them less likely to want to be worthless and nobody likes you."
active on social media. This can also affect the reputation of
 To tokenize this sentence, we could use a
the platform itself if it becomes known as a place where straightforward method of separating the text by
bullying is rampant. whitespace and punctuation marks.
3) Decreased engagement: Cyber-bullying can lead to
decreased engagement on social media, as individuals may  The tokenized sentence would look like this:
avoid posting or interacting with others out of fear of being  ["I", "hate", "you", "and", "wish", "you", "were",
targeted. This can have a negative impact on the platform as a "never", "born", ".", "You're", "worthless", "and",
whole, as it relies on user engagement to generate revenue. "nobody", "likes", "you", "."]

553 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 14, No. 7, 2023

Each element in the resulting list is a token that can be  Stop word removal is a valuable tool for detecting
processed and analyzed further using various NLP techniques, cyber-bullying and promoting a safer online
Fig. 2. environment.
 However, it is essential to note that stop-word removal
alone may not be enough to identify instances of cyber-
bullying accurately and that other techniques, such as
sentiment analysis and machine learning algorithms,
may be required.
3) Feature extraction techniques for cyber-bullying:
Feature extraction is a crucial step in natural language
processing (NLP) tasks such as cyber-bullying detection. Here
are some techniques for feature extraction from cyber-bullying
text data:
a) Bag of Words (BoW): It is a simple and effective
method to extract features from text data. It involves counting
the frequency of occurrence of each word in the document.
For instance, consider the following sentence: "You're such a
loser. Nobody likes you." The BoW representation of this
sentence would be: {'you': 2, 're': 1, 'such': 1, 'a': 1, 'loser': 1,
'nobody': 1, 'likes': 1}.
b) TF-IDF (Term Frequency-Inverse Document
Frequency): It is another technique that helps to extract
features from text data. It assigns a weight to each word based
on its frequency in the document and its frequency in the
entire corpus. For example, in the sentence "You're such a
loser. Nobody likes you," the word "you" appears twice in the
document but is likely to appear in many other documents too.
So, the weight assigned to "you" will be relatively low.
c) N-grams: N-grams are a sequence of N words in a
sentence. For example, a bigram of the sentence "You're such
a loser. Nobody likes you" would be "you're such," "such a,"
"a loser," "loser nobody," "nobody likes," and "likes you." N-
Fig. 2. Overall system architecture.
grams help capture the context of words in a sentence.
2) Stop-words removal: Stop words are commonly used in d) Word embeddings: Word embeddings are vector
a language but have little meaning and can be removed from representations of words that capture semantic and syntactic
text without affecting the overall message. Stop word removal relationships between them. They are learned using neural
can be used in cyber-bullying to filter out irrelevant or networks trained on large amounts of text data. Word2Vec and
GloVe are some examples of popular word embedding
offensive words from the text to identify and prevent bullying
techniques.
behavior.
Example:
 Here's an example of how to use stop-word removal in
the context of cyber-bullying:  Let's say you have a dataset containing cyber-bullying
text data, and you want to use these techniques to
 Assume a social media platform wants to look for extract features from it. Here is an example of how you
instances of cyber-bullying in user posts. can use these techniques to extract features:
 The platform could include a stop word filter that  Suppose you have a sentence in your dataset like this:
removes common words and phrases that are unlikely "You are ugly and nobody likes you."
to be used in cyber-bullying incidents, such as "the,"
"and," "is," "in," "a," "of," and "on."  BoW representation: {'you': 2, 'are': 1, 'ugly': 1, 'and': 1,
'nobody': 1, 'likes': 1}.
 For example, if a user writes, "I hate you, and I hope
you die," the stop word filter will remove the words  TF-IDF representation: {'you': 0.276, 'are': 0.276,
"I," "you," "and," "hope," and "die." The filtered text 'ugly': 0.385, 'and': 0.385, 'nobody': 0.385, 'likes':
would then be "hate," raising a red flag and prompting 0.385}.
a review by the platform's moderators.

554 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 14, No. 7, 2023

 N-gram representation: {('you', 'are'): 1, ('are', 'ugly'): 1, A. Convolutional Neural Networks (CNN) for Training on
('ugly', 'and'): 1, ('and', 'nobody'): 1, ('nobody', 'likes'): Images and Videos
1, ('likes', 'you'): 1}. CNNs are a type of deep learning model that are
 Word embeddings: [-0.456, 0.678, -0.234, 0.987, particularly effective at processing visual data, making them a
0.678, -0.567] (this is just an example of a vector popular choice for image and video classification tasks,
representation of the sentence using word embeddings, including cyber bullying detection. The basic architecture of a
and the values are random). CNN consists of several layers, including convolutional
layers, pooling layers, and fully connected layers. Each layer
V. ASPECT-BASED SENTIMENT ANALYSIS (ABSA) FOR performs a specific function, and the output of one layer is fed
CYBER-BULLYING as input to the next layer.
ABSA is a natural language processing technique used to The equations used to train a CNN for cyber bullying
identify and extract aspects or features of a given text and image and video classification involve the use of back-
determine their sentiment polarity (positive, negative, or propagation and gradient descent to update the weights and
neutral) [22]. In the context of cyber-bullying, ABSA can be biases of the network. The overall goal is to minimize the
used to identify the specific aspects or topics that are error between the predicted output and the actual output.
associated with negative or abusive comments, messages, or
posts. The general equation for computing the output of a
convolutional layer can be expressed as follows:
A mathematical model for ABSA in cyber-bullying
detection could be formulated as follows: (∑ ̂ ) (7)

Let D be a set of documents containing potentially abusive Where:

or negative content, and A be a set of aspects or topics that  ‘yi’ is the output of the ith neuron in the layer.
may be associated with cyber-bullying. Each document d ∈ D
can be represented as a set of sentences * + and  ‘f()’ is the activation function.
each sentence si can be further represented as a set of
words * + . Let ( ) be the probability of  ‘n’ is the number of input neurons.
word wi occurring in sentence si, and let ( ) be the  ‘wj’ is the weight connecting the jth input neuron to the
probability of sentence si occurring in document d. ith output neuron.
The sentiment polarity of each aspect a ∈ A can be  ‘xij’ is the activation of the jth input neuron at the ith
determined based on the sentiment scores of the words that are location of the receptive field.
associated with that aspect. Let S(a) be the sentiment score of
aspect a, which can be calculated as follows:  ‘bi’ is the bias term for the ith output neuron.
( ) ∑ ∈ ( ) ( ) (5) The pooling layer reduces the dimensionality of the input
by aggregating nearby activations. The most common pooling
Where ( )the polarity scores of is word (e.g., operation is max pooling, which selects the maximum value
+1 for positive, -1 for negative, 0 for neutral). from each local neighborhood of activations.
To detect cyber-bullying, we can use a threshold value T The fully connected layer takes the flattened output from
to determine whether a document d contains abusive or the previous layer and applies a matrix multiplication
negative content. Let B(d) be a binary variable that indicates operation to produce the final output. The equation for the
whether document d is abusive or not, where B(d) = 1 if d is fully connected layer can be expressed as:
abusive and B(d) = 0 otherwise. We can define B(d) as
follows: ( ) (8)
( ) * ∈ ( ) + (6) Where:
Where ∈ ( ) is the maximum sentiment score of y is the output vector.
all aspects in document d. The threshold value T can be f() is the activation function.
determined empirically based on the distribution of sentiment
scores in a training dataset of labeled cyber-bullying and non- W is the weight matrix.
cyber-bullying documents. x is the input vector.
Overall, the mathematical model for ABSA in cyber- b is the bias vector.
bullying detection involves identifying the aspects or topics
associated with cyber-bullying, calculating the sentiment During training, the weights and biases of the network are
scores of those aspects based on the sentiment polarity of the updated using the back-propagation algorithm. The gradient of
words associated with them, and using a threshold value to the loss function with respect to each weight and bias is
determine whether a document is abusive or not. computed, and the weights and biases are updated in the
opposite direction of the gradient to minimize the loss
function.

555 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 14, No. 7, 2023

The equations for back-propagation and gradient descent E. Image Classification using CNN
are as follows: Train a CNN on the image data to determine whether or
not each image contains cyber-bullying content.
(9)
Multiple convolutional and pooling layers are followed by
(10) fully connected layers and a softmax output layer in a CNN.
To improve the model's performance, employ data
(11) augmentation, dropout, and early stopping techniques.
F. Dataset Description
(12) The experiments use a Python programming language with
three datasets: Twitter, Images, and Videos dataset. The
Where: dataset is aged between 15-40 years from schools to job
‘L’ is the loss function. holders. Among these, 88% of data is analyzed as cyber-
bullying. These tweets contain more than 47656 with two
‘w’ and ‘b’ are the weights and biases of the network. attributes such as tweet_text and tweet_type. Python libraries
‘y’ is the output of the network. such as Keras, Pandas, and TensorFlow were used to analyze
the performance of the proposed model. Table I shows the
‘lr’ is the learning rate. various types of bullying text messages for training and testing
Overall, the use of CNNs with back-propagation and is given.
gradient descent provides an effective way to train models for Table II shows the types of bullying images that affects the
cyber bullying image and video classification tasks. human personally and mentally. These images are JPAG
B. Cropping of Images and Videos images with standard size. These images are classified based
on comments, captions, and topics.
Cropping is the removal of unwanted parts of an image or
video to focus on a specific area or subject. Table III shows the various types of videos belong to
different categories. Three types of bullying videos are present
Here are some typical image and video cropping for experimental analysis. These videos such as hate speech,
techniques: personal abuse and normal are shown in Table IV.
1) The rule of thirds: This technique entails dividing the
TABLE I. TYPES OF CYBER-BULLYING TEXT MESSAGES
image or video into thirds horizontally and vertically and then
positioning the subject along the intersections or lines. Types of Cyber-
Tweets
bullying
As a result, the composition is more balanced and visually
The girl who bullied you in high school but now wants
appealing. Age
to sell you Arbonne
2) Center crop: This technique involves cropping an
I said dont put north west in coffee fuck the diddy call
image or video to center the subject. It works well when the Ethnicity fifty i said there no to assassinate out the door of the
issue is the main focus and there is no distracting background. air port dumb niggers
Cropping an image or video to a specific aspect ratio, such as Don't call bitches females. That's mad disrespectful.
Gender
4:3 or 16:9, is an example of this technique. It comes in Bitches hate when you call them females.
handy when creating content for specific platforms or devices. @UmarMal And I'm not sure how you can yammer
3) Pan and zoom: This technique involves cropping an Religion about homelessness when Muslims are still murdering
people for apostacy and blasphemy.
image or video and animating it to simulate a camera pan or
Other type of @Eleoryth I sometimes envy those who don't have
zoom. It can add motion to the image or video or emphasize Cyber-bullying retarded parents
specific parts.
Not Cyber-bullying Rebecca Black Drops Out of School Due to Bullying
4) Content-aware crop: This technique uses software
tools to determine the best cropping based on the image or TABLE II. TYPES OF TWITTER TEXT DATASET
video content. It can be helpful when the subject is not in a
fixed position or when complex background content needs to Message Type Training Testing
be removed. Religion based bullying 3000 4997
C. Deep Neural Network (DNN) Age based bullying 3000 4992
DNN is specifically CNNs and recurrent neural networks Ethnicity 3000 4959
(RNNs), can be used to solve the problem of classifying Gender 3000 4948
cyber-bullying in images and videos (RNNs).
Other cyber-bullying 3000 4823
D. Data Gathering and Pre-processing Not cyber-bullying 3000 4937
Preprocess a large dataset of images and videos containing Total 18000 29656
Cyber-bullying content to extract features like color, texture,
shape, and motion.

556 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 14, No. 7, 2023

TABLE III. TYPES OF IMAGE DATASET Precision: This parameter gives the overall correct outputs
given by the proposed model.
Image Type Training Testing
Morphing Images 1500 1500 (13)
Personal Abuse 500 500
F1 Measure: It is the parameter that combines the recall
Adult 1k 1k and precision.
Total Messages 3k 3k
(14)
TABLE IV. VIDEOS DATASET
Accuracy: The overall accuracy of proposed model is
Video Type Training Testing measured as
Hate speech 25 25
(15)
Personal Abuse 25 25
Recall: This metric is mainly focused on reducing the
Normal Videos 10 10
false negatives.
Total Videos 60 60
(16)
G. Performance Metrics
A confusion matrix is an approach that analyzes the
performance of the proposed model. The classification of
images will specify the model performance on test data. The
confusion matrix mainly focused on two attributes such as
predicted and original values, see Fig. 3.
True Negative (TN): The predicted input is bullied and
actual input is also bullied.
True Positive (TP): The predicted input is not bullied and
actual value is not bullied.
False Positive (FP): The predicted input is bullied and
actual input is not bullied.
False Negative (FN): The predicted input is not-bullied
and actual input is bullied. Fig. 3. Confusion matrix.

TABLE V. COMPARATIVE PERFORMANCES OF EXISTING AND PROPOSED APPROACHES FOR ANALYSIS OF TWITTER DATA

Precision F1 Measure Accuracy Recall

Religion based bullying 0.68 0.68 0.68 0.68
Age based bullying 0.69 0.698 0.68 0.69
Ethnicity 0.67 0.68 0.68 0.67
2DCNN [23]
Gender 0.678 0.687 0.698 0.685
Other cyber-bullying 0.68 0.68 0.687 0.685
Not cyber-bullying 0.68 0.675 0.675 0.68
Religion based bullying 0.87 0.87 0.87 0.87
Age based bullying 0.886 0.885 0.876 0.868
Ethnicity 0.88 0.87 0.873 0.871
InceptionV3 [24]
Gender 0.878 0.876 0.879 0.873
Other cyber-bullying 0.875 0.878 0.897 0.867
Not cyber-bullying 0.884 0.878 0.876 0.875
Religion based bullying 0.965 0.956 0.966 0.964
Age based bullying 0.961 0.966 0.968 0.967
Ethnicity 0.976 0.975 0.978 0.967
Proposed Approach
Gender 0.956 0.967 0.976 0.956
Other cyber-bullying 0.976 0.976 0.968 0.976
Not cyber-bullying 0.978 0.972 0.976 0.967

557 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 14, No. 7, 2023

Precision F1 Measure Accuracy Recall

1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
Religion based bullying

Gender

Other cyber-bullying

Not cyber-bullying

Religion based bullying

Gender

Other cyber-bullying

Not cyber-bullying

Religion based bullying

Gender

Other cyber-bullying

Not cyber-bullying
Ethnicity

Ethnicity

Ethnicity
Age based bullying

Age based bullying

Age based bullying

2DCNN [23] InceptionV3 [24] Proposed Approach

Fig. 4. Comparative performances of existing and proposed approaches for analysis of Twitter data.

Table V shows the comparative results among the existing

Precision F1 Measure Accuracy Recall
2DCNN [23], InceptionV3 [24] and proposed approach. Fig. 4
also compares text based cyber-bullying models based on the
given data in twitter dataset. 0.9
0.8
Table VI, Table VII and Fig. 5, Fig. 6 shows a comparison 0.7
of image-based cyberbullying. MoSI is the existing model, 0.6
0.5
and Ensemble CNN is the proposed model. Ensemble CNN 0.4
classifies images and videos of cyber-bullying. They are 0.3
creating a diverse and representative dataset of cyberbullying 0.2
incidents, including various types such as harassment, hate 0.1
speech, or threats, and creating deep learning model 0
Adult

Adult
Morphing Images

Morphing Images
Personal Abuse

Normal

Personal Abuse

Normal
architecture suitable for detecting cyberbullying. The model
should be trained and optimized for high performance using
the collected dataset. They are addressing the issue of
imbalanced data, as cyber-bullying incidents are relatively rare
compared to non-cyber-bullying incidents. These techniques Motion from Static Images Ensemble CNN
need methods like oversampling, under sampling, and (MoSI)
generating synthetic data.
Fig. 5. Comparative performances of existing and proposed approaches for
TABLE VI. COMPARATIVE PERFORMANCES OF EXISTING AND PROPOSED analysis of image data.
APPROACHES FOR ANALYSIS OF IMAGES
TABLE VII. COMPARATIVE PERFORMANCES OF EXISTING AND PROPOSED
F1 APPROACHES FOR ANALYSIS OF VIDEOS
Types Precision Accuracy Recall
Measure
Morphing F1
0.78 0.78 0.78 0.78 Precision Accuracy Recall
Images Measure
Motion from Personal Hate
0.79 0.698 0.78 0.79 Motion 0.68 0.78 0.78 0.78
Static Images Abuse speech
(MoSI) from Static Personal
Adult 0.77 0.78 0.78 0.77 0.69 0.698 0.78 0.79
Images Abuse
Normal 0.77 0.78 0.78 0.78 (MoSI)[25]
Normal 0.67 0.78 0.78 0.77
Morphing
0.88 0.88 0.88 0.88 Hate
Images 0.80 0.81 0.81 0.81
speech
Personal Ensemble
0.89 0.898 0.88 0.89 Personal
Ensemble CNN Abuse CNN 0.80 0.821 0.82 0.81
Abuse
Adult 0.87 0.88 0.88 0.87
Normal 0.81 0.81 0.81 0.82
Normal 0.8856 0.879 0.8876 0.876

558 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 14, No. 7, 2023

[2] Alam, K.S., Bhowmik, S., Prosun, P.R.K.: Cyber-bullying detection: an

Precision F1 Measure Accuracy Recall ensemble based machine learning approach. IEEE: 2021 Third
International Conference on Intelligent Communication Technologies
and Virtual Mobile Networks (ICICV), pp. 710–715 (2021).
0.9
[3] S. Agrawal and A. Awekar, "Deep learning for detecting Cyber-bullying
across multiple social media platforms," in Advances in Information
0.8 Retrieval (Lecture Notes in Computer Science), vol. 10772, G. Pasi, B.
Piwowarski, L. Azzopardi, and A. Hanbury, Eds. Cham, Switzerland:
0.7 Springer, 2018, pp. 141–153.
[4] K. Reynolds, A. Kontostathis, and L. Edwards, "Using machine learning
0.6
to detect Cyber-bullying," in Proc. 10th Int. Conf. Mach. Learn. Appl.
Workshops (ICMLA), vol. 2, Dec. 2011, pp. 241–244,
0.5
doi:10.1109/ICMLA.2011.152.
0.4 [5] A. Muneer and S. M. Fati, "A comparative analysis of machine learning
techniques for Cyber-bullying detection on Twitter," Futur. Internet, vol.
0.3 12, no. 11, pp. 1–21, 2020, doi: 10.3390/fi12110187.
[6] N. Yuvaraj, K. Srihari, G. Dhiman, K. Somasundaram, A. Sharma, S.
0.2 Rajeskannan, M. Soni, G. S. Gaba, M. A. AlZain, and M. Masud,
"Nature-inspired-based approach for automated Cyber-bullying
0.1 classification on multimedia social networking," Math. Problems Eng.,
vol. 2021, pp. 1–12, Feb. 2021, doi: 10.1155/2021/6644652.
0 [7] N. Yuvaraj, V. Chang, B. Gobinathan, A. Pinagapani, S. Kannan, G.
Hate Personal Normal Hate Personal Normal Dhiman, and A. R. Rajan, "Automatic detection of Cyber-bullying using
speech Abuse speech Abuse multi-feature based artificial intelligence with deep decision tree
classification," Comput. Electr. Eng., vol. 92, Jun. 2021, Art. no.
Motion from Static Images Ensemble CNN 107186, doi: 10.1016/j.compeleceng.2021.107186.
(MoSI)
[8] Y. Zhang and A. Ramesh, "Fine-grained analysis of Cyber-bullying
using weakly-supervised topic models," in Proc. IEEE 5th Int. Conf.
Fig. 6. Comparative performances of existing and proposed approaches for Data Sci. Adv. Anal. (DSAA), Oct. 2018, pp. 504–513,
analysis of video data. doi:10.1109/DSAA.2018.00065.
[9] N. M. G. D. Purnamasari, M. A. Fauzi, Indriati, and L. S. Dewi, "Cyber-
VI. CONCLUSION bullying identification in Twitter using support vector machine and
information gain based feature selection," Indones. J. Electr. Eng.
Finally, an ensemble deep neural network can classify Comput. Sci.,vol. 18, no. 3, pp. 1494–1500, 2020, doi:
cyber-bullying on Twitter using text, images, and videos. The 10.11591/ijeecs.v18.i3.pp1494-1500.
ensemble method combines multiple models' outputs to [10] Zhao and K. Mao, "Cyber-bullying detection based on
produce a more accurate and robust prediction. An Aspect- semanticenhanced marginalized denoising auto-encoder," IEEE Trans.
based Sentiment Analysis (ABSA) for Cyber-bullying is Affect. Comput., vol. 8, no. 3, pp. 328–339, Jul. 2017,
introduced for text classification. This model learned the doi:10.1109/TAFFC.2016.2531682.
patterns and features that distinguish cyber-bullying tweets [11] X. Zhang, J. Tong, N. Vishwamitra, E. Whittaker, J. P. Mazer, R.
Kowalski, H. Hu, F. Luo, J. Macbeth, and E. Dillon, "Cyber-bullying
from non-cyber-bullying tweets using large datasets of labeled detection with a pronunciation based convolutional neural network," in
text. The text model's output can then be combined with the Proc. 15th IEEE Int. Conf. Mach. Learn. Appl. (ICMLA), Dec. 2016, pp.
results of the image and video models via a weighted voting 740–745, doi: 10.1109/ICMLA.2016.0132.
scheme. A deep neural network, such as a recurrent neural [12] Y. Zhang and A. Ramesh, "Fine-grained analysis of Cyber-bullying
network (RNN) or a convolutional neural network (CNN), can using weakly-supervised topic models," in Proc. IEEE 5th Int. Conf.
classify images and videos. These models trained on large Data Sci. Adv. Anal. (DSAA), Oct. 2018, pp. 504–513, doi:
10.1109/DSAA.2018.00065.
datasets of labeled images and videos to learn the features and
patterns that differentiate cyber-bullying from non-cyber- [13] R. R. Dalvi, S. B. Chavan, and A. Halbe, "Detecting a Twitter Cyber-
bullying using machine learning," in Proc. 4th Int. Conf. Intell. Comput.
bullying content. A weighted voting scheme is used for the Control Syst. (ICICCS), May 2020, pp. 297–301, doi:
output of the image and video models that can be combined 10.1109/ICICCS48265.2020.9120893.
with the output of the text model. Because different models [14] R. Zhao and K. Mao, "Cyber-bullying detection based on semantic-
can capture various aspects of the data, using an ensemble enhanced marginalized denoising auto-encoder," IEEE Trans. Affect.
deep neural network allows for a more accurate and robust Comput., vol. 8, no. 3, pp. 328–339, Jul. 2017, doi:
classification of cyber-bullying on Twitter. The ensemble 10.1109/TAFFC.2016.2531682.
model is better equipped to handle the complexity and [15] Luo, Y., Zhang, X., Hua, J., Shen, W.: Multi-featured Cyber-bullying
detection based on deep learning. IEEE: 2021 16th International
variability of cyber-bullying content on Twitter because their Conference on Computer Science & Education (ICCSE), pp. 746–751
outputs are combined. Overall, using an ensemble deep neural (2021).
network is a promising approach for addressing the problem [16] Adav, J., Kumar, D., Chauhan, D.: Cyber-bullying detection using pre-
of cyber-bullying on Twitter and can contribute to creating a trained bert model. IEEE: 2020 International Conference on Electronics
safer and more positive online environment. and Sustainable Communication Systems (ICESC), pp. 1096–1100
(2020).
REFERENCES [17] Ahmed, M.T., Rahman, M., Nur, S., Islam, A., Das, D.: Deployment of
machine learning and deep learning algorithms in detecting Cyber-
[1] Mahlangu, T., Tu, C.: Deep learning Cyber-bullying detection using
bullying in bangla and romanized bangla text: A comparative study.
stacked embbedings approach. IEEE: 2019 6th International Conference
IEEE: 2021 International Conference on Advances in Electrical,
on Soft Computing & Machine Intelligence (ISCMI), pp. 45–49 (2019).

559 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 14, No. 7, 2023

Computing, Communication and Sustainable Technologies (ICAECT), [22] Sahana, B., Sandhya, G., Tanuja, R., Ellur, S., Ajina, A.: Towards a
pp. 1–10 (2021). safer conversation space: Detection of toxic content in social media
[18] Iwendi, C., Srivastava, G., Khan, S., Maddikunta, P.K.R.: Cyber- (student consortium). In: 2020 IEEE Sixth International Conference on
bullying detection solutions based on deep learning architectures. Multimedia Big Data (BigMM), pp. 297–301 (2020).
Multimedia Systems, 1–14 (2020). [23] Kumari K, Singh JP, Dwivedi YK, Rana NP (2020) Towards Cyber-
[19] Aind, A.T., Ramnaney, A., Sethia, D.: Q-bully: a reinforcement learning bullying-free social media in smart cities: a unified multimodal
based Cyber-bullying detection framework. IEEE: 2020 International approach. Soft Comput 24(15):11059–11070
Conference for Emerging Technology (INCET), pp. 1–6 (2020). [24] Roy PK, Tripathy AK, Das TK, Gao X-Z. A framework for hate speech
[20] Ketsbaia, L., Issac, B., Chen, X.: Detection of hate tweets using machine detection using deep convolutional neural network. IEEE Access.
learning and deep learning. In: 2020 IEEE 19th International Conference 2020;8:204951–204962. doi: 10.1109/ACCESS.2020.3037073.
on Trust, Security and Privacy in Computing and Communications [25] ZHuang, S. Zhang, J. Jiang, M. Tang, R. Jin, and M. H. Ang, "Self-
(TrustCom), pp. 751–758 (2020). supervised motion learning from static images," in Proceedings of the
[21] Pradhan, A., Yatam, V.M., Bera, P.: Self-attention for Cyber-bullying ieee/cvf conference on computer vision and pattern recognition,
detection. In: 2020 IEEE International Conference on Cyber Situational Nashville, TN, USA, June 2021.
Awareness, Data Analytics and Assessment (CyberSA), pp. 1–6 (2020).
IEEE

560 | P a g e
www.ijacsa.thesai.org