Leveraging Deep Neural Networks for Real‑Time Detection of

Cyberbullying and Offensive Memes in Social Network

K. Jayasurya, T. Periyasamy, S. Prasanth, S. Sibhisaran and A. Ananya

Department of Artificial Intelligence and Machine Learning, M. Kumarasamy College of Engineering, Karur, Tamil Nadu,

India

Keywords: Social Media, Freedom of Speech, Hate Speech, Cyberbullying, Sentiment Analysis.

Abstract: Although connection to other people is possible through the use of social media, the result has been what the

social media users use to call as freedom of speech, which means that hate speech is also possible, and cyber

bullying. Free speech has come to mean liberty to post vengeful memes or to harass individuals or groups

based on their gender, colour, or religion. To locate such risky content, this study recommends a two-part

approach that involves natural language processing and optical character recognition. The model takes textual

and the visual input through deep neural recognition and analyses the sentiment of the text through a sentiment

recognition dictionary and a sentiment estimator. Negative content is filtered according to script and feedback

provided by the users of the application. When implemented on social media networks, these insights, which

are delivered in real-time, can significantly improve the fight against the use of technology to perpetrate abuse.

1 INTRODUCTION

But most important is – how open and available the

SM communication is. Open one or several accounts

in social networks and begin to share with other

people opinions, ideas and beliefs. But it has now

become a crucial facet of the current society because

modern society does not know it. As they said that all

those people who spread hate in one way or the other,

for example in politics, racists etc. are the ones who

are most likely to become either abusive or harassing

towards them in such a type of stage. Some ideas like

the use of politics or racism usually occur via the

various social media platforms. Unfortunately, such

purposes are increasingly often used for such actions

as blackmail, abuse, and even computer criminality.

By this it has become easier to know and interact with

other communities or organizations of like nature

through networks. Slightly above a third of these

network users are below 30 years of age. These

platforms are filled with a wealth of information to

enable researchers undertake comprehensive analysis

within targeted sub-topics of study. The process of

identifying cyber bullying is depicted by the

following the figure 1 below.

Figure 1: Social media types.

Dialectically, via social media, members of societies

worldwide can communicate, exchange information,

build personal relationships, and influence one

another (Mozafari et al.; Mosca et al.). However,

convenience brings risks, particularly among youth

under 30, who face threats like online blackmail,

cyberbullying, and emotional distress (Sabat et al.;

Gravano et al.). Social networking platforms facilitate

communication between individuals and

organizations while simultaneously generating vast

quantities of user-generated content (Alkomah &

Ma). Among the most commonly applied methods to

Jayasurya, K., Periyasamy, T., Prasanth, S., Sibhisaran, S. and Ananya, A.

Leveraging Deep Neural Networks for Real-Time Detection of Cyberbullying and Offensive Memes in Social Networ k.

DOI: 10.5220/0013908200004919

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 4, pages

69-74

ISBN: 978-989-758-777-1

analyze this content is sentiment analysis, widely

used in previous research (Fortuna et al.; Aluru et al.).

1.1 Background

Social media operates as both a communication and

an information-sharing platform due to its global

reach (Patil et al.). However, the same digital

connectedness allows harmful content like memes

containing hate speech or provocative humor, and

cyberbullying messages to spread rapidly (Toraman

et al.; Antypas & Camacho-Collados). Cyberbullying

involves deliberate acts of harassment or threats using

digital tools, while offensive memes often relay

derogatory and hateful messages with even greater

virality (Sabat et al.; Zhou et al.). Such content affects

users’ mental health, undermines platform standards,

and threatens the credibility of social media

businesses (Velankar et al.; Akuma et al.). Given the

volume of textual content online, it is infeasible for

human moderators to review every post, even with

filtering rules (Mullah & Zainon; Rabiul Awal et al.).

This calls for real-time, automated detection systems

to mitigate the spread of harmful material (Gravano

et al.; Malik et al.).

2 LITERATURE REVIEW

This research applies a Deep Convolutional Neural

Network (DCNN) to detect hate speech and offers a

multilingual study using 16 datasets across 9

languages (Roy et al.; Aluru et al.). While hate speech

has long been studied in humanities, its exploration in

computer science remains nascent (Alkomah & Ma;

Fortuna et al.). One key proposal is DeepHate, a deep

learning model leveraging high-dimensional text

representations for robust detection (Cao et al.).

Another approach, BiCHAT, combines BiLSTM,

CNN, and hierarchical attention for nuanced

detection (Khan et al. 2022a), while the AngryBERT

model jointly learns target and emotion contexts

(Rabiul Awal et al.). A BERT-based transfer learning

technique has also demonstrated success in detecting

hate speech from social media platforms (Mozafari et

al.).

The recently proposed HCovBi-Caps model

integrates convolutional layers, BiGRU, and capsule

networks to improve hate speech detection,

particularly considering its abstract nature (Khan et

al. 2022b). Further, methods like TF-IDF and Bag of

Words have been evaluated in combination with

machine learning models for hate classification on

live tweets (Akuma et al.). The cross-domain

transferability of hate speech detection models is

examined by Toraman et al., while Antypas &

Camacho-Collados evaluate robustness across

datasets. Meanwhile, context-based methods (Mosca

et al.) and sentiment-sharing models (Zhou et al.)

have enhanced semantic understanding in hate speech

systems. Moreover, offensive meme detection,

focusing on pixel-level content, pushes hate

moderation into multimodal dimensions (Sabat et al.).

3 EXISTING METHODOLOGIES

And even if individuals are communicating

asynchronously over the Internet, social networking

as a stylish and, most importantly, unobtrusive way

for people to show their respect to each other’s

individual opinions and beliefs is already a part of

everyday life. Today, the grown man who expresses

hatred through politics, misogyny, and racial

prejudice has harassed and mistreated other people.

Social networks are also gaining popularity among

users for online tyrant, including blackmail. Myspace

is a common social networking site through which

people can comfortably join one or many societies

that they find interesting Due to the sheer amount of

information that exists in SNSs, researchers have

been able to conduct extensive research in a number

of fields. Emotion analysis is one of the areas of

research that receive a lot of attention and primarily

use data from social media. Recommending filters

arrange the offered content list in relation to the

similarity between the desired item content and the

end user’s interests; whereas, collaborative filters

offer items based upon comparison with other like-

minded individuals. Since content-based filtering

processes are more or less focused on text-based

documents, Filtering, which categorizes the mass

incoming texts into relevant and irrelevant, must truly

be regarded as a single tag classification. Multi-label

text categorization is an advanced form of filter that

will categorize the communications partly by topics

and categories. The classifier generates automatically

according with the learning need from the set of

previously classified sample in content-based

filtering paradigm from the machine learning (ML)

framework, identified that Bag of Words (BoW)

method perform better when compare with complex

text representation method that may have better

statistical features but lower semantics.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

4 PROPOSED

METHODOLOGIES

Another common and well-known tool of information

exchange share and communication involving

personal data is considered to be an online social

network (OSN). But one of the significant challenges

for OSNs is to provide the users with control over

their privacy and stopping the broadcasting of

unsuitable content. This work provides a solution to

the problem by allowing the users of OSN to decide

on the messages that appear on their walls. This is

done through allowing the users to set up their own

specific filtering rules of their choosing in a highly

flexible rule-based system. Moreover, there is

utilizing of the soft classifier founded on the option of

the machine learning for the messages’ automatic

recognition and content filtering. Every short

message is analyzed by four Deep Learning (DL) text

classification algorithms, which categorize the

message into one or more groups based on its content.

A dataset of the categorized words is developed to

check for any words that should not be in any

particular category.If there is any inappropriate

language in the correspondence, it will be filtered to

remove it from the Blacklists. Moreover, since the

wall message uses material technique, no immoral

word will be written on the wall of the user.

Figure 2: Framework.

Attributes such as text message content, users’ buddy

connections and other writer related characteristics

are also considered by a system, utilizing blacklists

for filtering contextless messages. The criteria of the

proposed structure include the application of deep

learning in helping the user define FRs used as an

understanding of filter rules, enhancement of fit rules

in the considered domain, and the expansion of the set

of attributes to be analyzed initially by the classifier.

context messages. The suggested framework

incorporates deep learning techniques to assist users

in defining Filtering Rules (FRs), improve fit rules

within the domain under consideration, and broaden

the collection of attributes initially examined during

the classification process. Figure 2 shows the

framework.

4.1 Framework

Social Networking Service or SNS is a particular type

of online platform that makes it possible to keep links

with people who have common interests, activities,

backgrounds, or acquaintance for all people,

regardless if it is casual or business related. As this

point grows increasingly flexible, it has been very

difficult to define social media. Social network is

human relations that are in which people share,

exchange and pass values, knowledge and

experience. Such interactions can be enhanced by

creating a graphical user interface where most of user

communication and necessary graphical and visual

triggers can be undertaken. Both user and admin

interfaces may be designed using this module. We are

providing users to interact with the system, respond

to invitations, post photos with friends, and upload

Leveraging Deep Neural Networks for Real-Time Detection of Cyberbullying and Offensive Memes in Social Network

meme images. They can then be forwarded to the

admin page to determine and analyze.

4.2 Words Extraction

Social networking has recently become more and

more an integral part of everyday online activities as

social apps and websites are taking social networking

forward as an important medium of communication.

Today social networking sites allow people to do a lot

of things that stimulate active participation such as

leave user comments and share multimedia. But

businesses know social media can help grow and

increase visibility and use social media for marketing,

brand promotion, connecting with customers as well

as making new alliances. Online social networks that

allow for free interaction and commentary are now

open discussion, however, they need content

management systems. By extracting text from

photographs through optical character recognition (or

text recognition), or OCR, content management,

filtering, and analysis on these platforms will be

improved.

4.2.1 Optical Character Recognition

• Extraction of Character boundaries from

Image,

• Building a Bag of Visual Words (BOW)

framework in remembering the Character

images,

• Loading trained Model,

• Consolidating predictions of characters

The text may consist of bigrams, multigram, or

unigrams. Feedback from social media users is

gathered using this component. Links, text, and brief

messages are just a few of the various formats that

comments might take. After being reviewed,

comments are sent to the server page.

4.2.2 Text Mining

Developing an effective deep learning classifier often

emphasizes identifying and extracting key features

that define and characterize the data. The typical

process in text mining consists of these steps:

• Exclude commonly used words that do not

add value to the analysis.

• Reduce words to their root forms to

standardize variations.

• Remove symbols, punctuation, and other

non-alphanumeric characters.

• Identify and isolate sentences that convey

the most relevant information.

• Simplify or replace lengthy phrases with

more appropriate or concise alternatives.

By taking these steps the text is pre-processed well

for the deep learning models to effectively classify

and analyse the text. By taking these steps the text is

pre-processed well for the deep learning models to

effectively classify and analyse the text.

4.3 Classification

In this module, we built an automatic mechanism

which we called Filtered Wall (FW) to filter

unsolicited communications coming from OSN user

walls. The three-tiered architecture supports OSN

services. The initial layer generally attempts to

provide the basic OSN capabilities (relations and

profiles). Additionally, some OSNs offer a further

layer allowing us to deal with external Social

Network Application (SNA). In the last, an additional

layer may be required for supporting SNA's graphical

user interfaces (GUIs). This thesis centres around the

development of a robust backpropagation neural

network (BPNN) that aims to extract and select a set

of discriminative and characteristic features. Text

classification interfaces between the constraints that

are specified, and the constraints that will be applied

in the process. In the approach, it is assumed that the

easiest to remove (likely best) is "neutral" sentences,

which do not contribute significantly to the

classification task. Finally, the remaining ‘non-

neutral’ sentences are assigned into the appropriate

‘classes’ of interest. This approach eliminates the

burden of the work and allows us to process the data

in a more organized fashion since, working from the

VADER point of view, we suggest a hierarchical two-

level method.

4.3.1 Developing the Deep Learning Model

The process works in developing a system to capture

and remove any offensive content from user

generated text. Here's the proposed workflow for

creating a deep learning-based classifier:

• First, we have to start with a collection of a

different kind of data or comments. This

includes both examples of offensive (hateful)

and non-offensive (neutral) language in this

data.

• Remove any character, symbol, or irrelevant

information that might interfere with analysis

in cleaned collected text.

• Therefore, deep learning such as the VADER

sentiment analysis algorithm or similar natural

language processing (NLP) algorithms is used

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

to train a model to differentiate between

hateful and non-hateful language.

• With the processed data feed into the model,

you want to use and tweaking the model’s

parameters according to the labelled examples

so the model can learn how to correctly

classify hate speech.

• After training, run the model inside a system

that can rapidly and correctly identify

offensive language of new user inputs to keep

things positive.

• Watchlists can also be used by the system to

keep track of offensive terms or expressions;

These lists auto flag content depending on the

relationship and context with the user, their

message. This screening can be tailored to the

specific needs of the domain so that the

filtering procedure is precise and appropriate

in any situation.

By the use of deep learning methods to classify

objectionable language, our solution ensures that

objectionable content is still effectively controlled

while increasing user experience.

4.4 Filtering Guidelines

Filtering should allow users to place restrictions on

content creators to suit particular circumstances. Such

guidelines can be based on conditions by

characteristics represented in user profiles. For

example, guidelines can be tailored according to the

likes of managing them for young creators, people

who espouse a certain religious or political view or

those deemed to have started in a particular subject

and lack experience. By applying filters to profile

attribute, such as work-related information, it can be

achieved. It can also use content-based parameters to

identify, track and manage communication. Take the

case of automatically banning people who keep

posting negative reviews above a certain level — five

times, let’s say. Additionally, such actions should be

notified to the users through those mobile devices.

4.5 Alert System

The BL overseeing system should be able to locate

individuals in the BL and know when customer

loyalty goals have been complete. The set of BL

specific rules guiding this process is intended to

increase system adaptability and threshold values are

key to these rules. Having set the thresholds, the

server defines acceptable levels of activity. This value

can then be used by users to do actions like allowing

or disallowing the posting of individuals that often

comment in a critical way. In addition, the system

should also send real time user notifications via

smartphone notifications to maintain transparency

and updated interfacing.

5 CONCLUSIONS

In this work, the hybrid model for harmful memes and

cyberbullying in the social media environment is

investigated, and we use the VADER sentiment

analysis algorithm along with natural language

processing to accomplish this. VADER helps the

model recognize the subtle linguistic patterns often

found in cyberbullying and is a proven effective

processor of social media material leading to a

reputable sentiment measure. Once OCR is added to

image analysis, the combination of VADER can

reliably detect offensive multimedia messages such

as abusive memes containing grotesque text

incorporated in images. The system’s adaptability

with respect to filtering options is also increased

through management of BLs. This is the first stage of

a bigger project. The early encouraging results with

the classification technique have motivated us to

work on other projects in order to improve

classification quality. The DL soft classifier is used to

filter out unwanted signals in this system. Using BL

increases the filtering system versatility. We will

design a system that incorporates more factors in

determining whether or not to introduce a user into

the BL. With a flexible language, the strong rule layer

generates Filter Rules (FRs) that constrain the system

to not display some data on their walls. The use in

FRs is to utilize user profiles and relationships.

REFERENCES

Akuma, Stephen, Tyosar Lubem, and Isaac Terngu Adom.

"Comparing Bag of Words and TF-IDF with different

models for hate speech detection from live tweets."

International Journal of Information Technology 14.7

(2022): 3629-3635.

Alkomah, Fatimah, and Xiaogang Ma. "A literature review

of textual hate speech detection methods and datasets."

Information 13.6 (2022): 273.

Aluru, Sai Saketh, et al. "Deep learning models for

multilingual hate speech detection." arXiv preprint

arXiv:2004.06465 (2020).

Antypas, Dimosthenis, and Jose Camacho-Collados.

"Robust hate speech detection in social media: A cross-

dataset empirical evaluation." arXiv preprint

arXiv:2307.01680 (2023).

Leveraging Deep Neural Networks for Real-Time Detection of Cyberbullying and Offensive Memes in Social Network

Cao, Rui, Roy Ka-Wei Lee, and Tuan-Anh Hoang.

"DeepHate: Hate speech detection via multi-faceted

text representations." Proceedings of the 12th ACM

Conference on Web Science. 2020

Fortuna, Paula, et al. "Directions for NLP Practices Applied

to Online Hate Speech Detection." Proceedings of the

2022 Conference on Empirical Methods in Natural

Language Processing. 2022.

Gravano, Agustín, et al. "Assessing the Impact of

Contextual Information in Hate Speech Detection."

IEEE Access, vol. 11, pp. 30575-30590, 2023, doi:

10.1109/ACCESS. 2023.3258973. (2023).

Khan, Shakir, et al. "HCovBi-caps: hate speech detection

using convolutional and Bi-directional gated recurrent

unit with Capsule network." IEEE Access 10 (2022):

7881-7894.

Khan, Shakir, et al. "BiCHAT: BiLSTM with deep CNN

and hierarchical attention for hate speech detection."

Journal of King Saud University-Computer and

Information Sciences 34.7 (2022): 4335-4344.

Malik, Jitendra Singh, Guansong Pang, and Anton van den

Hengel. "Deep learning for hate speech detection: a

comparative study." arXiv preprint arXiv:2202.09517

(2022).

Mosca, Edoardo, Maximilian Wich, and Georg Groh.

"Understanding and interpreting the impact of user

context in hate speech detection." Proceedings of the

Ninth International Workshop on Natural Language

Processing for Social Media. 2021.

Mozafari, Marzieh, Reza Farahbakhsh, and Noel Crespi. "A

BERT-based transfer learning approach for hate speech

detection in online social media." Complex Networks

and Their Applications VIII: Volume 1 Proceedings of

the Eighth International Conference on Complex

Networks and Their Applications COMPLEX

NETWORKS 2019 8. Springer International

Publishing, 2020.

Mullah, Nanlir Sallau, and Wan Mohd Nazmee Wan

Zainon. "Advances in machine learning algorithms for

hate speech detection in social media: a review." IEEE

Access 9 (2021): 88364-88376.

Patil, Hrushikesh, Abhishek Velankar, and Raviraj Joshi.

"L3cube-mahahate: A tweet-based marathi hate speech

detection dataset and bert models." Proceedings of the

Third Workshop on Threat, Aggression and 2022.

Rabiul Awal, Md, et al. "AngryBERT: Joint Learning

Target and Emotion for Hate Speech Detection." arXiv

e-prints (2021): arXiv-2103.

Roy, Pradeep Kumar, et al. "A framework for hate speech

detection using deep convolutional neural network."

IEEE Access 8 (2020): 204951-204962.

Sabat, Benet Oriol, Cristian Canton Ferrer, and Xavier

Giro-i-Nieto. "Hate speech in pixels: Detection of

offensive memes towards automatic moderation."

arXiv preprint arXiv:1910.02334 (2019).

Toraman, Cagri, Furkan Şahinuç, and Eyup Halit Yilmaz.

"Large-scale hate speech detection with cross-domain

transfer." arXiv preprint arXiv:2203.01111 (2022).

Velankar, Abhishek, Hrushikesh Patil, and Raviraj Joshi.

"A review of challenges in machine learning based

automated hate speech detection." arXiv preprint

arXiv:2209.05294 (2022).

Zhou, Xianbing, et al. "Hate speech detection based on

sentiment knowledge sharing." Proceedings of the 59th

Annual Meeting of the Association for Computational

Linguistics and the 11th International Joint Conference

on Natural Language Processing (Volume 1: Long

Papers). 2021.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES