Safe Browsing for Kids Under Parental Supervision Using Machine

Learning

Pilli Suneetha, Vaartha Sai Sruthi, Vemula Ghnana Sri Sai Lakshmi, Sirigiri Malavika

and Koyye Vijaya

Department of Information Technology, Sagi Rama Krishnam Raju Engineering College,

Bhimavaram 534202, Andhra Pradesh, India

Keywords: Safe Browsing, Machine Learning, Deep Learning, Content Filtering, Parental Controls, Natural Language

Processing (NLP), Image Classification, Reinforcement Learning.

Abstract: The internet has become a vital resource for children's education and pleasure due to the quick digitization

of modern life. They are, nevertheless, exposed to potentially damaging materials, such as offensive, violent,

or graphic material. In order to establish a kid-friendly online environment under parental supervision, this

project presents Safe Browsing for Kids, a comprehensive solution that combines Machine Learning (ML)

and Deep Learning (DL) approaches. The system uses cutting-edge techniques such as Support Vector

Machines (SVM) for website classification, Convolutional Neural Networks (CNN) for picture filtering, and

Natural Language Processing (NLP) for textual content analysis. The system is further adjusted to changing

browsing behaviours via a dynamic Reinforcement Learning (RL) method. The first steps in the process

include gathering browsing data (URLs, metadata, and images), preprocessing it, and applying ML/DL

models to categorize content into safe or risky groups. Through an intuitive dashboard, a real-time monitoring

system not only filters harmful websites but also gives parents notifications and comprehensive surfing

reports. The findings show that hazardous content may be filtered with high accuracy, limiting exposure to

unsuitable content while protecting kids' online experiences. Additionally, over time, adaptive learning

improves the accuracy of the system. This project effectively establishes a digital environment for kids that is

trust-based, safe, and educational. It gives parents cutting-edge tools to safeguard their children online by

striking a balance between privacy and supervision. To sum up, the system guarantees a safer online

experience, encouraging responsible online behavior while tackling the ever-changing issues of web safety in

the contemporary period.

1 INTRODUCTION

K., Keshwala. (2024). The internet has become a

commonplace part of modern life, providing a wealth

of opportunities for learning, communication, and

entertainment. Children, in particular, have easy

access to the internet through a variety of devices,

including laptops, tablets, and smartphones.

However, while the internet presents opportunities for

positive experiences, it also exposes young users to

harmful and inappropriate content, including

violence, explicit material, and cyberbullying.

Bandaru, et al, 2024 These risks present serious

challenges for parents who want to strike a balance

between their children's safety and encouraging their

independence and curiosity online. These issues have

not been adequately addressed by conventional

content filtering techniques, such as blockers based

on keywords. Static filters frequently miss context-

dependent content, coded language, and changing

online behaviours. Additionally, they frequently

either under block, allowing harmful content to pass

through, or over block, limiting access to safe and

instructive sites. This insufficiency calls for a more

clever and flexible strategy to protect kids' internet

safety. A comprehensive solution is offered by this

project: "Safe Browsing for Kids." Chien, et al, 2022

The system's goal is to establish a safe and instructive

online environment under parental supervision by

combining cutting-edge Machine Learning (ML) and

Deep Learning (DL) techniques. Real-time content

filtering, adjustable parental controls, and adaptive

Suneetha, P., Sruthi, V. S., Lakshmi, V. G. S. S., Malavika, S. and Vijaya, K.

Safe Browsing for Kids Under Parental Supervision Using Machine Learning.

DOI: 10.5220/0013924300004919

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 5, pages

149-160

ISBN: 978-989-758-777-1

149

learning are some of the system's primary features.

Xianjun, et al, 2022 The system overcomes the

drawbacks of conventional solutions by utilizing

methods like Natural Language Processing (NLP) for

textual analysis, Convolutional Neural Networks

(CNN) for image and video filtering, and

Reinforcement Learning (RL) for continual

improvement.

Bandaru, et al, 2024 The suggested system's

capacity for dynamic content analysis and

classification is one of its main advantages. For

example, NLP techniques like transformers (e.g.,

BERT) allow the detection of hate speech and

destructive language, even when it is presented using

coded terminology or slang. Akintunde, et al, 2024

CNN models can also recognize violent or graphic

imagery, which guarantees that visual content is

properly filtered. Bandaru, et al, 2020 By learning

from kids' browsing patterns and adjusting filters

appropriately, Reinforcement Learning further

increases the system's adaptability. The system's

design places a strong emphasis on parental

involvement. Parents may create filters based on age,

interests, and educational needs, as well as receive

real-time warnings and comprehensive information

on their child's browsing activities, all through an

intuitive dashboard. This gives parents the ability to

actively supervise their kids' internet activities while

upholding privacy and trust. Jennyphar, et al, 2022 In

contrast to conventional methods, which frequently

function as "black boxes," this system places a strong

emphasis on openness and cooperation, promoting a

more positive relationship between parents and their

kids online. Another crucial component of the system

is its real-time monitoring capabilities. The system

uses a lightweight browser plugin to continuously

analyze text, video, and website URLs in order to

detect and filter harmful information. Without

interfering with their online experience, this

guarantees that kids are exposed to as little hazardous

content as possible. In order to promote healthy

internet practices, the system also suggests

educational and kid-friendly websites.

Bandaru, et al, 2023 This endeavour is important

for reasons other than only households. Maintaining

a secure online environment is becoming a social

necessity as kids depend more and more on digital

platforms for socialization and education. This

project’s innovative use of AI-driven technologies

addresses this need, providing a scalable and adaptive

solution that can evolve alongside the internet’s

dynamic landscape. To sum up, the "Safe Browsing

for Kids" initiative is a ground-breaking strategy for

protecting kids online. Milind, et al, 2021 It provides

a strong, flexible, and approachable answer to

contemporary digital problems by utilizing state-of-

the-art ML and DL algorithms. In addition to

shielding kids from dangerous material, this system

gives parents the resources they need to foster their

kids' curiosity and education in a secure online

setting.

1.1 Role of Machine Learning and

Deep Learning in Ensuring Safe

Browsing for Children

A parental security control tool was proposed by

Milind et al. to guarantee children have safer internet

experiences. The technology dynamically classifies

webpages according to their content using a machine

learning-based methodology. It successfully filters

dangerous or inappropriate websites that could

expose kids to explicit content or other internet

hazards by using this classification. Additionally,

parents can modify access controls and keep an eye

on their child's browsing history. This tool

emphasizes how crucial real-time filtering is to

guaranteeing prompt action against harmful content.

Although the strategy shows a lot of promise, it only

focuses on fundamental machine learning methods,

leaving space for integration with more sophisticated

models, such as deep learning, to increase scalability

and adaptability in dynamic online contexts.

A controlled internet browsing technique that uses

machine learning for webpage classification was

presented by Arup et al. Their approach allows

parents to impose content limits that are customized

to meet their child's needs by classifying websites

according to user profiles and pre-established

regulations. Additionally, the tool has a feedback

feature that gives parents information about the

websites they visit. The technology adjusts to user

preferences by incorporating machine learning,

guaranteeing a safer online experience. However,

because deep learning and reinforcement learning are

not used for real-time categorization, the system lacks

dynamic adaptability. Additionally, its ability to

handle contemporary online hazards like offensive

images or videos is limited by the lack of multimedia

content filtering. In order to overcome these

constraints, the study establishes a strong basis for

future improvements.

In order to predict the hazards to children's online

safety, such as exposure to abuse, cyberbullying,

sexting, and stress, Rumel et al. created a machine

learning model. This approach places a strong

emphasis on proactively identifying children who are

at risk and emphasizes how important parental

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

150

participation is in reducing these risks. The system

alerts parents in advance by examining user behavior

and spotting patterns of susceptibility. The method

does not filter harmful information or block access to

dangerous websites in real time, despite its heavy

emphasis on risk prediction. There are gaps in

immediate internet safety since it depends on parental

response following risk identification. The study

emphasizes that in order to provide complete kid

safety solutions, predictive models must be integrated

with active monitoring and filtering systems.

Patel, D., et al. reaffirmed in a different study the

significance of machine learning in tackling threats to

children's online safety and privacy. Potential

vulnerabilities include exposure to hazardous content,

cyberbullying, and privacy breaches are predicted by

the model. It places a strong emphasis on parental

supervision and active participation in their children's

internet activities. However, this work lacks

substantial implementation breakthroughs and mostly

aligns with the team's previous conclusions. Real-

time content filtering and automated techniques to

instantly ban harmful content are not included in the

theoretical model. Building on this framework, future

studies may integrate dynamic filtering technologies

with predictive analytics to offer a more proactive and

approachable method of ensuring children's online

safety.

The Kids Safe Search Classification Model was

presented by Deepshikha et al. and combines a Neural

Network Classifier to filter unsuitable content, PCA

for feature selection, and Modified Entropy word

weighting. The purpose of this strategy is to improve

children's surfing experiences without the need for

continuous adult supervision. While the incorporation

of PCA increases feature extraction for greater

accuracy, machine learning guarantees effective

content classification based on safety parameters.

Notwithstanding its advantages, the model can only

be used for textual content classification; it cannot be

used for multimedia filtering. Furthermore, its use in

dynamic online situations is diminished by its lack of

real-time implementation. Zhao, et al, 2014 This

study serves as a foundation for future systems that

incorporate multimedia content screening and real-

time monitoring.

A framework for voice analysis and machine

learning was presented by Anonymous et al. with the

goal of identifying and removing dangerous internet

content. The approach emphasizes the value of

human moderation for complicated content

interpretation while concentrating on identifying

language that is aggressive, explicit, or otherwise

harmful. With this hybrid method, basic filtering

duties are handled by machine learning models, while

complex and unclear cases are handled by human

oversight. Although the system is excellent in many

ways, its automation and scalability are limited by its

heavy reliance on human interaction. Moreover, real-

time adaptation and multimedia filtering are not

adequately covered in the study. Lee, et al, 1999 In

addition to highlighting the possibility of integrating

AI and human expertise, this study emphasizes the

necessity of additional automation in order to

properly manage changing online dangers.

Expanding on their earlier research, Anonymous

et al. investigated machine learning's potential to

shield kids from dangerous internet content. In order

to comprehend complicated, ambiguous

circumstances that automated systems might

overlook, their architecture incorporates human

moderation. This method does not apply to visual

content like pictures or movies; instead, it

concentrates on censoring explicit text and language.

Scalability and reaction time issues arise from the

dependence on manual intervention, particularly in

dynamic online situations. Notwithstanding its

drawbacks, the framework emphasizes how crucial

human supervision is to guaranteeing correct content

interpretation, opening the door for more

sophisticated systems that strike a balance between

automation and human knowledge for complete child

protection.

A safe online browsing system with content

screening and parental control modules was proposed

by Zhao et al. Parents can use the system to restrict

websites according to particular safety standards and

modify theme libraries. Although it successfully

handles fundamental content filtering requirements,

the system's lack of machine learning and deep

learning capabilities restricts its capacity to adjust to

changing internet threats. It is less efficient at

managing complicated content or dynamic browsing

patterns since it relies too heavily on static filtering

techniques. Notwithstanding these shortcomings, the

study offers insightful information about the

significance of adaptable parental controls and gives

a basis for incorporating cutting-edge AI-driven

strategies to improve online child safety.

Using trust lists, URL approval processes, and

user profiles, Cary et al. presented a technique for

parental internet monitoring. Using a whitelist of

trusted URLs and pre-approving websites, this

approach enables parents to control their children's

internet usage. However, the approach is

inappropriate for today's dynamic internet contexts

because it does not make use of machine learning or

real-time adaptability. The method is static and cannot

Safe Browsing for Kids Under Parental Supervision Using Machine Learning

151

handle dangers that depend on context or multimedia

content. Despite being out of date, the study

emphasizes the need of user-centric customization in

parental control systems and provides a basis for

more advanced solutions that integrate real-time

filtering and artificial intelligence.

Alghowinem, et al, (2018). CNN, RNN, and

OpenCV technologies were used by deep learning-

based architecture for secure browsing. When children

or teenagers visit the system, it automatically hides

inappropriate content and uses facial recognition to

determine the user's age. This architecture exhibits

dynamic flexibility to user profiles and powerful

multimedia filtering capabilities. However, its limited

applicability to non-visual content and dependence on

facial recognition present privacy problems.

Notwithstanding these difficulties, the work shows

how deep learning may be used to build reliable and

adaptable systems for secure browsing, opening the

door for the integration of thorough content filtering

systems with improved privacy protection. Table 1

show the Literature survey.

Table 1. Literature Survey.

Author Method/Technique Key Findings Gaps/Limitations

Proposed

Enhancements

Milind et al.

(2021)

Machine Learning

(Dynamic Website

Classification)

Real-time

filtering of

harmful content

and parental

monitoring

Limited to basic ML

methods; lacks deep

learning capabilities

Integrate

advanced deep

learning models

for scalability and

adaptabilit

Arup et al.

(2017)

Content Classification

using User Profiles

Allows

customized

content limits and

provides feedback

to parents

No multimedia filtering;

lacks real-time

adaptability

Incorporate

multimedia

filtering and

reinforcement

learning for

dynamic

ada

tabilit

Rumel et al.

(2023)

Machine Learning Risk

Prediction Model

Predicts

risks like

cyberbullying,

sexting, and

online abuse

Focuses only on

prediction; lacks real-time

filtering capabilities

Combine

prediction with

active filtering

systems for

comprehensive

rotection

Patel et al.

(2016)

Kids Safe Search

Classification Model

Efficient

textual content

filtering using

PCA and Neural

Networks

Limited to textual

data; lacks real-time and

multimedia filtering

Expand to

multimedia

content analysis

and enable real-

time monitoring

Anonymous

(2023)

ML and Speech Analysis

Framework

Detects

harmful content

using audio and

textual analysis

Heavy reliance on

human moderation; lacks

automation

Improve

automation with

advanced models

for scalability and

real-time

adaptation

Anonymous

(2023)

Hybrid ML Framework

with Human Moderation

Balances ML

filtering and

human oversight

for accuracy

Limited scalability and no

support for multimedia

filtering

Enhance

automation and

include

multimedia

capabilities for

comprehensive

rotection

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

152

Zhao et al.

(2014)

Parental Control with

Theme Libraries

Customizable

content filtering

rules for safer

rowsing

Static filtering techniques;

no ML integration

Use ML/DL for

adaptive filtering

and

ersonalization

Cary et al.

(1999)

User Profiles and URL

Whitelist

Provides basic

parental control

using static trust

lists

Outdated method; no real-

time adaptability or ML

Introduce real-

time AI-driven

filtering

mechanisms

Anonymous

(2022)

Deep Learning with

CNN, RNN, and

OpenCV

Dynamic age

detection and

multimedia

filtering

abilities

Limited to visual content;

potential privacy concerns

Broaden to

include non-visual

filtering and

enhance privacy

measures

Sanders et al.

(2016)

Technology-Related

Parenting Strategies

Highlights the

role of parenting

in managing

screen time

No technical or automated

solutions for filtering

Develop AI-based

tools for proactive

and adaptive

filterin

solutions

3 OBJECTIVE AND

METHODOLOGY

3.1 Objective’s

The main goal of this research is to create a machine

learning-based system that offers kids a secure and

instructive online environment. This system seeks to:

• Dynamically filter and block offensive material,

such as obscene pictures, videos, and text.

• Allow parents to receive real-time monitoring and

notifications.

• Give parents programmable controls so they can

impose limits according to their child's interests

and age.

• In order to adjust to new online hazards and

behaviours, make use of reinforcement learning.

The solution fills in the deficiencies seen in previous

methods, such as poor real-time adaptation, lack of

multimedia filtering, and inadequate scalability.

Input: Text data, Image dataset, States (S), Actions (A), Rewards (R)

Output: Classified labels (Safe/Unsafe), Optimal policy

1. Text Classification:

a. Tokenize the text into words or phrases.

b. Compute TF-IDF scores for the tokens.

c. Train an SVM classifier using labelled textual data.

d. Predict labels (Safe/Unsafe) for unseen textual data.

2. Image Classification:

a. Preprocess images by resizing and normalizing.

b. Pass the images through convolutional layers to extract features.

c. Apply ReLU activation to introduce non-linearity.

d. Perform max pooling to reduce dimensionality.

e. Flatten the feature maps and classify the images using a fully connected layer.

3. Reinforcement Learning (Dynamic Rule Optimization):

a. Initialize Q-values for all state-action pairs.

b. For each episode:

i. Select an action using the epsilon-greedy policy.

ii. Observe the reward and the next state.

iii. Update Q-values using the formula:

Q(s, a) ← Q(s, a) + α * [r + γ * max(Q(s', a')) - Q(s, a)]

c. Derive the optimal policy from the updated Q-values.

Safe Browsing for Kids Under Parental Supervision Using Machine Learning

153

4. Combine Results:

a. Aggregate text and image classifications from Steps 1 and 2.

b. Use the optimized policy from Step 3 to dynamically adjust filtering rules.

c. Out

ut the final classification labels

(

Safe/Unsafe

)

and the u

dated

olic

3.2 Techniques & Methodology

This solution combines several machine learning

(ML) and deep learning (DL) approaches to guarantee

efficient content filtering and secure browsing for

kids. By tackling certain tasks like content analysis,

image classification, and dynamic learning, each

algorithm makes a distinct contribution to the total

functioning. These methods are explained in depth

below, complete with pseudocode, mathematical

formulae, and illustrations.

National Language Processing (NLP): Analyzing

textual information to determine if it is safe or

harmful is the main goal of NLP. In feature extraction,

the TF-IDF (Term Frequency-Inverse Document

Frequency) formula is essential. The frequency of a

term t in a document d is measured by the Term

Frequency (TF) parameter, which is computed as

shown in Equation (1):

TF(

(

t,d

)



,

∑







,



∈

(1)

Where 𝑓

,

is the count of term t in the document d,

and

∑

𝑓





,





∈

represents the total term in the

document?

The Inverse Document (IDF) parameter evaluates the

rarity of a term across the corpus D, Calculated and

shown in Equation (2):

IDF

(

T,D

)

=log







|

∈:∈

|



(2)

Where

𝐷

is the total number of Document, and



𝑑∈𝐷:𝑡∈𝑑



is the number of documents

containing t. The TF-IDF score is the obtained as

shown in Equation (3):

𝑇𝐹 − 𝐼𝐷𝐹

(

𝑡,𝑑,𝐷

)

=𝑇𝐹

(

𝑡,𝑑

)

∙𝐼𝐷𝐹(𝑡,𝐷) (3)

SVM Derivation: By optimizing the margin between

classes, a Support Vector Machine (SVM) classifier

uses these attributes to distinguish between safe and

harmful content. The decision boundary of the SVM

is explained by as shown in Equation (4):

𝑚𝑖𝑛

,





𝑤



𝑠𝑢𝑏𝑗𝑒𝑐𝑡 𝑡𝑜 𝑦



(

𝑤∙𝑥



+𝑏

)

≥1𝑓𝑜𝑟 𝑎𝑙𝑙 𝑖 (4)

Where 𝑤 is the weight vector, 𝑥 is the feature vector,

and b is the bias. This classifier ensures high accuracy

in textual content analysis as shown in Equation (5 &

6).

𝑚𝑎𝑥



∑

𝜆



−





∑∑

𝜆



𝜆



𝑦



𝑦



𝑥



∙𝑥

















(5)

𝑓

(

𝑥

)

=𝑠𝑖𝑔𝑛 (

∑

𝜆



𝑦



(

𝑥



∙𝑥

)

+𝑏)





(6)

Convolution Neural Network (CNNs): CNNs

process image and videos by extracting features

through convolutions layers. The Convolution

operation calculates the weighted sum of a Kernel k

applied to the input image 𝑥 , producing a feature map

as shown in Equation (7):

ℎ



∑∑

𝑥

()()∙

𝑘



+𝑏









(7)

Where M and N are the dimensions of the kernel

𝑥

()()∙

𝑘



is the pixel intensity 𝑘



is the

kernel weight, and b is the bias term as shown in

Equation (8).

ℎ





=𝑚𝑎𝑥

(,)∈

ℎ

()()

(8)

This guarantees that only activations that are positive

are sent to layers that follow. By choosing the highest

value within a specified window, pooling layers like

max pooling lower the dimensionality of the feature

maps:

The fully connected layers aggregate features from

convolution layers for classification. The CNN is

optimized using a Cross- Entropy loss function as

shown in Equation (9).

𝐿=−

∑

𝑦



log (𝑦



)







(9)

Where 𝑦



is the true label and 𝑦



is the Predicted

probability?

The model's performance is determined by crucial

parameters such as activation thresholds, pooling

window (P), kernel size (M, N), and activation

thresholds, which guarantee precise visual content

filtering.

Reinforcement Learning RL: RL employs Markov

Decision Processes (MDP) to dynamically filtering

rules. An MDP consists of states S, action A,

transition probabilities 𝑃

(

𝑠



𝑠,𝑎

)

, rewards 𝑅

(

𝑠,𝑎

)

and a discount factor 𝛾. The value of a state-action

pair is updated using the Q-learning Algorithm as

shown in Equation (10).

𝑉

(

𝑠

)

=𝑚𝑎𝑥



[𝑅

(

𝑠,𝑎

)

+𝛾

∑

𝑃

(

𝑠



𝑠,𝑎

)

𝑉(𝑠



)]





(10)

𝑄

(

𝑠,𝑎

)

←𝑄

(

𝑠,𝑎

)

+𝛼[𝑟+𝛾𝑚𝑎𝑥





𝑄

(

𝑠



,𝑎



)

−

𝑄

(

𝑠,𝑎

)

] (11)

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

154

The reward system penalizes exposure to harmful

content while rewarding accurate classifications. The

policy π, which is formed from Q-values, converges

to the best filtering rules over iterations. Important

variables that balance immediate and long-term

rewards, such as learning rate (𝛼) and discount factor

(𝛾), guarantee strong adaptability. RL continuously

improves the filtering process by combining state-

action evaluation and dynamic decision-making,

which increases the system's reactivity to emerging

threats as shown in Equation (11).

Figure 2. Innovative System Architecture for Safe Browsing.

A multi-layered structure for guaranteeing children's

safe browsing is shown in the system architecture

diagram that is supplied. Through the Parental

Dashboard, which offers real-time updates and

activity reports, the User Interaction Layer enables

parents to keep an eye on and manage the system.

After processing material, the Browser in the

Application Layer uses a material Analyzer to

recognize text and multimedia inputs before sending

the information to the Filtering Engine for

categorization. Three essential modules are integrated

into the Decision and Analysis Layer the

Reinforcement Learning (RL) Module for dynamic

rule optimization, the CNN Module for spotting

dangerous images or videos, and the Natural

Language Processing (NLP) Module for textual

content analysis. Together, these modules provide

adaptive, real-time content filtering. Activity logs and

filtering rules are kept in the Data Management Layer

and are controlled by the Policy Manager. It allows

filtering criteria to be updated continuously based on

RL outputs. To guarantee strong online safety

measures, the overall architecture exhibits scalability,

adaptability, and a parent-friendly design as shown In

Figure 2.

4 RESULT AND VALIDATION

The outcomes demonstrate how well the system

filters and analyzes text, images, and multimedia

content in real time to create a secure browsing

Safe Browsing for Kids Under Parental Supervision Using Machine Learning

155

experience. The system's strengths are illustrated by

key performance measures like accuracy, latency,

adaptability, user experience, and scalability. With

high precision and recall rates, the NLP and CNN

modules are excellent at identifying unsuitable

information. By keeping latency low, the system

guarantees real-time filtering without interfering with

user experience. Filtering rules are dynamically

optimized using Reinforcement Learning, which

gradually increases accuracy and flexibility. The

dashboard's clear reports and easy-to-use controls are

validated by user feedback. Additionally, even with

high traffic, the system functions dependably under a

range of workloads with little deterioration.

4.1 Content Classification Accuracy

The above table highlights the performance metrics

of the NLP and CNN modules. The NLP module gets

good accuracy and F1-scores, indicating its

usefulness in recognizing dangerous textual content.

With a 92% accuracy rate and low false positive and

negative rates, the CNN module—which is intended

to identify images and videos also does well. These

metrics demonstrate how well the system works to

reduce pointless blocking (false positives) and make

sure hazardous content is not missed (false negatives)

as shown in Table 2 and Graphical Representation in

Figure 3.

Table 2: Performance Metrics of Content Classification

Modules.

Module

Latency

(ms)

Target

Latency

(ms)

Status

Text

Analysis

(NLP)

150 < 200 Achieved

Image/Video

Analysis

200 < 300 Achieved

Overall

System

Response

400 < 500 Achieved

Figure 3. Graphical Representation of Content

Classification Accuracy.

4.2 Latency and Real-Time

Performance

The latency of several modules is assessed in this

table. The goal of less than 200 ms is met by the NLP

module, which processes textual content with an

average latency of 150 ms. Likewise, the CNN

module meets its goal of analyzing visual content in

200 ms. Real-time performance appropriate for

browsing situations is ensured by the system's overall

response time, which includes both analysis and

decision-making phases, being less than 500ms.

These outcomes show that the system can quickly

filter information without compromising user

experience as shown in Table 3 and Graphical

Representation in Figure 4.

Table 3. Latency and Real-Time Performance Metrics.

Metric

Text

Analysis

(NLP)

Image/Video

Analysis (CNN)

Accuracy

(%)

95 92

Precision

(

)

96 93

Recall (%) 94 91

F1-Score

(%)

95 92

False

Positive Rate

(%)

2 3

False

Negative

Rate

(

)

3 4

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

156

Figure 4: Graphical Representation of Latency and Real-

Time Performance.

4.3 Reinforcement Learning

Adaptability

The flexibility of the reinforcement learning module

is demonstrated in this table. The accuracy of the

system's safe content filtering and unsafe material

blocking improved by 7% after ten learning episodes.

These outcomes demonstrate how the system can

dynamically learn and improve its filtering rules,

increasing its efficacy over time as shown in Table 4

and Graphical Representation in Figure 5. The rules'

convergence within ten episodes shows effective

policy adaptation, guaranteeing that there are few

delays until better performance is attained.

Table 4. Adaptability Metrics of Reinforcement Learning

Metric Rating (1-5)

User Feedback

(%)

Dashboard

Usabilit

4.5 90

Report Clarity 4.7 94

Control

Customization

Satisfaction

4.6 92

Figure 5: Graphical Representation of Reinforcement

Learning Adaptability.

4.4 User Experience Metrics

High levels of system satisfaction are indicated by

user experience indicators gathered from parent

questionnaires. Parents appreciated the dashboard's

accessible and user-friendly design, giving it a

usability rating of 4.5 out of 5. At 4.7/5, report clarity

received an even better grade, demonstrating the

system's capacity to provide insightful and thorough

information. A satisfaction rating of 4.6/5 was given

to the flexibility to personalize controls as shown in

Table 5 and Graphical Representation in Figure 6.

Ninety percent of parents gave excellent feedback

overall, highlighting how easy it was to use and how

well it enabled them to keep an eye on and control

their kids' internet use.

Table 5: User Experience Evaluation Metrics

Metric

Initial

Accuracy

(%)

Post-

Learning

Accuracy

(%)

Improvement

(%)

Safe Content

Filtering

88 95 +7

Unsafe

Content

Blocking

85 92 +7

Rule

Convergence

Time

episodes

Figure 6: Graphical Representation of User Experience

Metrics.

4.5 Scalability and Robustness

The system's scalability under various traffic loads

was assessed. The system exhibited little to no

performance degradation under light and moderate

Safe Browsing for Kids Under Parental Supervision Using Machine Learning

157

loads. The system's stability and scalability were

demonstrated when it maintained functionality with

only a 5% decrease even under situations of high

traffic (more than 1,000 requests per second). These

outcomes demonstrate that the system can efficiently

manage a range of workloads without sacrificing its

filtering or usability as shown in Table 6 and

Graphical Representation in Figure 7.

Table 6. System Scalability and Robustness Metrics.

Metric

Load

(Requests/Second)

Performance

Degradation

(%)

Light

Load (<

500

requests)

Degradation

Moderate

Load

(500-

1000

)

3 Minimal

Heavy

Load (>

1000)

5 Acceptable

Figure 7: Graphical Representation of Scalability and

Robustness.

5 CONCLUSIONS

The suggested solution successfully integrates

cutting-edge machine learning (ML) and deep

learning (DL) approaches to address the difficulties

of guaranteeing a secure surfing environment for

kids. The system exhibits strong performance in

filtering unsuitable textual and visual content with

high content classification accuracy, low latency for

real-time filtering, and dynamic adaptability through

reinforcement learning. The parental dashboard's

usability is confirmed by user feedback, which also

supports the system's function of enabling parents to

keep an eye on and control their kids' internet activity.

The system is a dependable solution for real-world

applications since the scalability and robustness

metrics attest to its capacity to manage a variety of

workloads while preserving efficiency.

Future research will concentrate on expanding the

system's functionality. The accuracy of textual

analysis will increase with the use of sophisticated

transformers, like GPT models, for improved

contextual comprehension of ambiguous content.

Visual content filtering will be strengthened by

adding multi-modal content analysis and deepfake

detection to the CNN module. Adaptability will be

strengthened by enhancing reinforcement learning

models for quicker convergence and more accurate

policy updates. The system will also be more secure

and inclusive if ethical issues are addressed, such as

minimizing bias in screening and guaranteeing

privacy compliance. Furthermore, the system's

usefulness and accessibility will be improved by

adding multilingual content support and connecting it

with mobile apps. With these developments, the

system will keep developing into a complete, flexible,

and scalable safe browsing solution, greatly

enhancing kids' online safety.

ACKNOWLEDGMENTS

Authors would like to thank all of the people and

organizations that helped them finish this research in

an indirect way. There was no outside funding or

financial assistance available for this investigation.

The study process was much enhanced by the insights

and conversations with peers and colleagues, which

offered insightful viewpoints. The writers are also

grateful for the helpful criticism they got while

preparing the manuscript, which enabled them to

improve it. The authors alone have contributed to this

study; no outside funding or institutional support has

been obtained.

REFERENCES

K., Keshwala. (2024). 11. A Comprehensive Review-

Building A Secure Social Media Environment for Kids-

Automated Content Filtering with Biometric Feedback.

International journal of innovative research in

computer science & technology, doi:

10.55524/ijircst.2024.12.4.4

Bandaru, V. N. R., Kaligotla, V. G. S., Varma, U. D. S. P.,

Prasadaraju, K., & Sugumaran, S. (2024, July). A

Enhancing Data Security Solutions for Smart Energy

Systems in IoT-Enabled Cloud Computing

Environments through Lightweight Cryptographic

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

158

Techniques. In IOP Conference Series: Earth and

Environmental Science (Vol. 1375, No. 1, p. 012003).

IOP Publishing.

Chien, Trong, Nguyen., Giang, Hoang, Nguyen., Long,

Khac, Pham., Anh, Nguyen., D., V., Nguyen., Son,

Ngo., Anh, Ngoc, Bui. (2022). 13. A Deep Learning

Based Application for Recognition and Preventing

Sensitive Image. doi: 10.1145/3556223.3556239

Xianjun, Meng., Shaomei, Li., Muhammad, Mohsin,

Malik., Qasim, Umer. (2022). 14. Machine-Learning-

Based Suitability Prediction for Mobile Applications

for Kids. Sustainability, doi: 10.3390/su141912400

Bandaru, V. N. R., Sumalatha, M., Rafee, S. M., Prasadraju,

K., & Lakshmi, M. S. (2024). Enhancing Privacy

Measures in Healthcare within Cyber-Physical Systems

through Cryptographic Solutions. EAI Endorsed

Transactions on Scalable Information Systems.

Akintunde, Nelson, Oshodi., Mojeed, Omotayo, Adelodun.,

Evangel, Chinyere, Anyanwu., Nkoyo, Lynn, Majebi.

(2024). 16. Combining parental controls and

educational programs to enhance child safety online

effectively. International journal of applied research in

social sciences, doi: 10.51594/ijarss.v6i9.1592

Bandaru, V. N. R., Kiruthika, S. U., Rajasekaran, G., &

Lakshmanan, M. (2020, December). Device aware

VOD Services with Bicubic Interpolation Algorithm on

Cloud. In 2020 IEEE 4th Conference on Information &

Communication Technology (CICT) (pp. 1-5). IEEE.

Jennyphar, Kavikairiua., Fungai, Bhunu, Shava. (2022). 18.

Algorithm to Impact Children’s Online Behaviour and

Raise their Cyber Security

Awareness. doi: 10.1109/icABCD54961.2022.985601

7 (2022). 19. Algorithm to Impact Children’s Online

Behaviour and Raise their Cyber Security

Awareness. doi: 10.1109/icabcd54961.2022.9856017

Bandaru, V. N. R., & Visalakshi, P. (2023, August). EEMS-

Examining the Environment of the Job Metaverse

Scheduling for Data Security. In International

Conference on Cognitive Computing and Cyber

Physical Systems (pp. 245-253). Cham: Springer

Nature Switzerland.

Milind, K., Dwivedi, V., Sanyal, A., Bhatt, P., & Koshariya,

R. (2021). Parental security control: A tool for

monitoring and securing children's online activities.

Proceedings of the ACM SIGMIS Conference.

https://doi.org/10.1145/3474124.3474196

Bhattacharya, A., Jun, J., & Wu, J. (2017). Method and

system to enable controlled safe internet browsing.

International Journal of Internet Security Research.

Rumel, M. S., Rahman, P. M., & Forhad, R. M. (2023).

Applying a machine learning model to forecast the risks

to children's online privacy and security. Proceedings of

the IEEE ISACC Conference 2023.

https://doi.org/10.1109/ISACC56298.2023.10084054

Patel, D., & Singh, P. K. (2016). Kids safe search

classification model. Proceedings of the IEEE CESYS

Conference 2016. https://doi.org/10.1109/CESYS.201

6.7914186

Anonymous. (2023). Machine learning and speech analysis

framework for protecting children against harmful

online content. Proceedings of the IEEE

ICEARS Conference 2023. https://doi.org/10.1109/IC

EARS56392.2023.10085565

Zhao, X., Zhang, K., Guo, L., & Wang, Y. (2014). Secure

web browsing system based on parental personalized

recommendation control. Journal of Internet Security

and Applications.

Lee, C., Bates, B. J., Cragun, P. R., & Day, P. R. (1999).

Method and computer program product for

implementing parental supervision for internet

browsing. Journal of Internet Control Systems.

Anonymous. (2022). Ensure safe internet for children and

teenagers using deep learning. Proceedings of the IEEE

DASA Conference 2022. https://doi.org/10.1109/dasa5

4658.2022.9765035

Alghowinem, S. (2018). A safer YouTube Kids: An extra

layer of content filtering using automated multimodal

analysis. Advances in Intelligent Systems and

Computing, Springer Verlag, 294-308.

Sanders, W., Parent, J., Forehand, R., & Breslend, N. L.

(2016). The roles of general and technology-related

parenting in managing youth screen time. Journal of

Family Psychology, 30(5), 641–646.

Sharifa, Alghowinem. (2018). 21. A Safer YouTube Kids:

An Extra Layer of Content Filtering Using Automated

Multimodal Analysis. doi: 10.1007/978-3-030-01054-

6_21

Christopher, Frye., Ilya, Feige. (2019). 22. Parenting: Safe

Reinforcement Learning from Human Input. arXiv:

Artificial Intelligence, Mike, Sullivan. (2003). 23.

Safety Monitor: How to Protect Your Kids Online.

Parry, Aftab. (1999). 24. The Parent's Guide to Protecting

Your Children in Cyberspace.

Nancy, E., Willard. (2007). 25. Cyber-Safe Kids, Cyber-

Savvy Teens: Helping Young People Learn To Use the

Internet Safely and Responsibly.

J., Bolton. (1999). 26. Keeping children safe on-line.

Multimedia information & technology, Rashid, Tahir.,

Faizan, Ahmed., Hammas, Saeed., Shiza, Ali., Fareed,

Zaffar., Christo, Wilson. (2019). 27. Bringing the kid

back into YouTube kids: detecting inappropriate

content on video streaming platforms. doi: 10.1145/3

341161.3342913

Chet, Thaker. (2011). 28. Tools for mobile safety for

children.

Saeed, Alqahtani., Wael, M., S., Yafooz., Abdullah,

Alsaeedi., Liyakathunisa, Syed., Reyadh, Alluhaibi.

(2023). 29. Children’s Safety on YouTube: A

Systematic Review. Applied Sciences, doi:

10.3390/app13064044

Tadikonda, Bala, Venkata, Naga, Abhya, Dattu., Pamarthi,

Bharath, Prabhakar., Davuluri, HemaLatha, Chowdary.,

Oleti, Dolly, Sumanta., G., Navya, Sree. (2024). 30.

Safe Browse Guardian. International Research Journal

on Advanced Science Hub, doi:

10.47392/irjash.2024.031

Xianjun, Meng., Shaomei, Li., Muhammad, Mohsin,

Malik., Qasim, Umer. (2022). 31. Machine-Learning-

Based Suitability Prediction for Mobile Applications

for Kids. Sustainability, doi: 10.3390/su141912400

Safe Browsing for Kids Under Parental Supervision Using Machine Learning

159

A., Chiwanza., Fungayi, D., Mukoko., B., Mupini. (2024).

32. A Web Crawling and NLP-Powered Model for

Filtering Inappropriate Content for Primary School

Learners' Online Research. International journal of

innovative science and research technology, doi: 10.3

8124/ijisrt/ijisrt24jul1083

Ta, Vinh, Thong. (2024). 33. A safety risk assessment

framework for children's online safety based on a novel

safety weakness assessment approach. arXiv.org, doi:

10.48550/arxiv.2401.14713

Dunja, Mladeni. (2000). 34. Machine learning for better

Web browsing. (2022). 35. Machine Learning for Web

Proxy Analytics. doi: 10.4018/978-1-6684-6291-

1.ch045 (2023). 36. Safeguarding your Data from

Malicious URLs using Machine Learning.

International Journal for Science Technology and

Engineering, doi: 10.22214/ijraset.2023.50907

Ahmad, El, Bakri., Abdullah, Yehia., Murtazza, Ali., Reem,

Osama., Zeyad, Adel., Omar, Gamal., Sherine, Nagy,

Saleh. (2024). 37. The Eye: An AI-Powered Video

Streaming Platform to Protect Children from

Inappropriate Content. doi: 10.1109/niles63360.2024.

1075322

S., Yazhini. (2023). 38. Child Digital Monitoring and

Controlling System. doi: 10.1109/ICAAIC56838.2023

.101415

Mark, Maldonado., Ayad, Barsoum. (2019). 39. Machine

Learning for Web Proxy Analytics. doi:

10.4018/IJCRE.2019070103

K., Gowsic., S, Siranjeevi., Nataliia, Shmatko., K., Swathi.

(2024). 40. Web spoofing defense empowering users

with phishcatcher's machine learning. ShodhKosh

Journal of Visual and Performing Arts, doi:

10.29121/shodhkosh. v5.i3.2024.2713

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

160