Evolution of Machine Learning Applications in IoT Security: A

Critical Analysis and Future Perspectives

Vaishali N. Rane and Arunkumar

Department of Computer Science and Engineering, Vel Tech Rangarajan Dr.Sagunthala R& D Institute of Science and

Technology, Avadi, Chennai, Tamil Nadu, India

Keywords: Internet of Things, Machine Learning, IoT Security, DDoS, Adversarial Attacks, Artificial Intelligence.

Abstract: The continuously advancing field of Internet of Things (IoT) has given rise to technologies that are crucial to

protect IoT systems from various attacks. Conventional security methods have their own bottlenecks and fail

to effectively deal with the security of IoT systems. This review explores the existing Artificial Intelligence

and Machine Learning (ML) based approaches for IoT security. Covering researches mainly done in 2023-

24, we have discovered that AI-enabled security solutions have better accuracy in detecting threats, more than

90%. On the other hand, they keep a check on the computational cost to prevent any cost overruns. Our key

takeaways include the integration of multiple ML algorithms as a hybrid system, adversarial attack handling

mechanisms and other techniques to handle targeted attacks. However, challenges persist, including resource

constraints, ease of deployment, adaptability, robustness etc. Our review addresses these challenges and

provide direction for future research along with the aim to offer valuable insights from relevant researches

aimed at enhancing IoT security through Machine Learning Techniques.

1 INTRODUCTION

The rapidly growing use of IoT devices in our daily

lives has completely changed how we interact with

such systems across homes, offices, and other

spheres. While offering numerous advantages to users

worldwide, the concern for security and data privacy

needs be addressed properly. Our review highlights

the bottlenecks of traditional IoT security systems and

in accordance with the research by (Ali & Rani,

2024), where they have stated that current systems are

not sufficient to protect against sophisticated and

more advanced attacks targeting IoT network.

The shift towards Machine Learning based

solutions is rapidly being preferred over traditional

methods as a solution for security challenges in this

domain. Unlike traditional approaches which are

based on static-rules and pre-defined conditions, ML

introduces dynamic and adaptable systems which can

protect against threats which learning from the data

continuously to keep itself updated with latest threats.

This milestone is described in the paper by (Saumya

et al.,2024) as a significant point in the building of

IoT security architectures. Table 1 shows the

Projected IoT connections worldwide.

The prime advantage of using machine learning

and AI in IoT security is its ability to deal with big

data and understand hidden patterns that normal

systems fail to uncover. This allows for better

protection in IoT environments where standard

protocols aren’t that much efficient in detecting

advanced cyberattacks. However, integrating AI/ML

in legacy IoT systems has its own set of challenges.

Some of them are resource constraints, low-power

operations, communication protocols, and

interoperability with various IoT devices. (Omarov et

al.,2024) have described these challenges in depth in

their research where they indicated that traditional

systems fail to maintain a balance between security

and simplicity of algorithm and how Machine

Learning based algorithms are better.

Rapid progress has been seen in past few years in

the field of Machine Learning in IoT. Some use cases

like real-time threat detection and prevention systems

have achieved accuracy as high as 97%. Research by

(Alomiri & AlShehri, 2024) describes how delay in

identification of threat can cause serious data

breaches and compromises in security.

Although research in this domain is going on, the

practical implementation of theoretical models and

systems has several challenges. One of the primary

Rane, V. N. and Arunkumar,

Evolution of Machine Learning Applications in IoT Security: A Critical Analysis and Future Perspectives.

DOI: 10.5220/0013866600004919

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 1, pages

409-414

ISBN: 978-989-758-777-1

409

Table 1: Projected IoT Connections Worldwide.

Year

IoT Connections

(Min, in

illions)

IoT Connections

(Max, in billions)

YoY Growth

(%) (Min)

YoY Growth

(%) (Max)

Key Observations

2022 13.8 13.8 - - Baseline Year

2023 15.9 16.6 15.2% 20.3%

Expansion of 5G and smart

devices

2024 18.0 18.8 13.2% 13.3%

Increased industrial IoT

ado

tion

2025 20.1 20.1 11.7% 6.9% Growth in smart city projects

2026 22.4 22.4 11.4% 11.4%

IoT penetration in healthcare

and AI

2027 24.7 24.7 10.3% 10.3%

Rise of edge computing and

6G research

2028 27.1 27.1 9.7% 9.7%

Autonomous vehicles and

AIoT sur

2029 29.6 29.6 9.2% 9.2%

Mass adoption of industrial

IoT

2030 32.1 40.0 8.5% 21.6%

IoT becoming mainstream in

dail

life

2031 34.6 34.6 7.8% -13.5% AI-driven IoT innovation

2032 37.1 37.1 7.2% 7.2% Global IoT regulations evolve

2033 39.6 39.6 6.7% 6.7%

IoT surpasses 39 billion

devices

challenges is adversarial attacks which manipulate

the training process and introduce errors in the

algorithm which makes it helpless against some

targeted attacks. (Harbi et al., 2024) have described

the risk of such attacks and how they disrupt the

machine learning model. Moreover, they have

highlighted the need for careful optimization of ML

models so they can perform efficiently under limited

resource conditions as well.

Another area of application, the Industrial Internet

of Things or IIoT has benefitted from the provision of

AI/ML based security systems. Adaptive models

allow autonomous handling of attacks, minimizing

human intervention. These systems can detect various

cyber threats including both known and new,

improving real-time monitoring and mitigation of

threats. (Wankhade et al., 2024) have described the

advancements in the field of IIoT and security

measures focusing mainly on precision, reliability

and real-time functioning.

Current trends mainly focus on the development

of hybrid machine learning models, combining more

than one algorithm to enhance it while overcoming

the limitations of one single algorithm. For instance,

combination of supervised and unsupervised

Machine Learning Algorithms has resulted in

increased threat detection accuracy and reduced false

positives as evaluated by (Iqbal et al., 2024).

Additionally, Pareto-optimal models have also been

discussed.

Significance of the Study. Understanding the role of

ML in IoT security and its impact, are crucial for

designing and developing advanced security

solutions. As cyber threats are getting advanced day

by day, so as the need to have adaptive and intelligent

security systems. This review summarizes the current

state of ML in IoT security highlighting various

techniques, major trends and key inferences which

can be derived from the study alone. The limitations,

challenges and future directions are clearly discussed

to help the researchers infer and conclude. Moreover,

it acts as a fundamental base, by providing the

directions of future research which can focus on

safety and reliability of IoT Environments and

optimization of complex ML algorithms to work

better with IoT units.

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

410

2 CURRENT STATE OF ML IN

IoT SECURITY

2.1 Intrusion Detection Systems

Intrusion detection systems (IDS) have emerged as

powerful tools for identifying and classifying threats

in IoT ecosystems. Their job is to detect any anomaly

in the system or network and alert the user before the

threat actually affects the system. Recent studies

show that Random Forest has displayed the accuracy

of 92.72% in binary classification of threats and

92.40% in multiclass classification tasks. Therefore,

it can be considered as a reliable choice for threat

detection systems. This review and comparison have

been done by (Saumya et al., 2024).

Figure 1: Advantages and disadvantages of Hybrid

approaches in IDS.

Similarly, other algorithms like Support Vector

Machines (SVM) have shown notable accuracy in

recall metrics, which makes them highly preferable

for detecting variety of tasks that too with greater

precision.

The performance of IDS can be further improved

using hybrid approaches. (Iqbal et al., 2024) have

experimented that integration of Supervised and

Unsupervised machine learning algorithms have

improved the detection accuracy and adaptability of

system to detect new and unknown threats, which the

traditional systems fail to identify. Advantages and

Disadvantages of Hybrid approaches in IDS is given

in figure 1.

2.2 Feature Selection and Performance

Optimization

Feature Engineering is crucial when we have mixed

and heterogeneous data to work upon. Since ML

model’s performance is determined by the quality of

training data, having features in bulk require Feature

selection and optimizing the process of feature

selection is a key requirement to identify which

weight vector is most important. According to (Htwe

et al., 2024), combining one-hot encoding with

Pearson correlation can achieve high accuracy while

maintaining computational load. This helps in

applications where we have limited resources like

computation, storage and network.

Figure 2: Enhancing ML Models in IoT security.

Pareto-optimal machine learning work best when

you have multiple performance objectives. (Wu et al.,

2024) highlights Gradient Boosting, Random Forest

and Decision Trees as some of the highly effective

machine learning approaches for building an anomaly

detection model. They primarily deal with the

detection accuracy, execution time and memory

usage and are mainly used in resource-constrained

environments. Enhancing ML Models in IoT Security

is shown in figure 2.

2.3 Real-Time Threat Detection

Another key feature of ML-based models is Threat

detection in real-time. Some systems have resulted

accuracy of over 97% in identification of network

threats (Alomiri & AlShehri, 2024). Also, in IIoT

systems, adaptive models have demonstrated high

accuracy and better threat response, by monitoring

data streams and based on that pattern, detect

anomaly in data and behavior. (Wankhade et al.,

2024) have explained how these models continuously

upgrade their threat detection capabilities, thus

providing robust security in dynamic environments.

Evolution of Machine Learning Applications in IoT Security: A Critical Analysis and Future Perspectives

411

Figure 4 depicts the Real-time threat detection in

IIoT.

Figure 3: Real-time threat detection in IoT.

2.4 DDoS Attack Detection and

Mitigation

Distributed Denial of Service (DDoS) attacks pose a

significant threat to network-based IoT systems. It

works by disrupting the availability of services and

data to various nodes. This is done by tampering data

nodes at network level. Recent studies have down that

Machine Learning based techniques have better

performance in detection and mitigation of DDoS

attacks (table 2).

Table 2: ML approaches for DDoS attack detection in IoT.

ML Technique

Detection

Rate

False Positive

Rate

Key Features Reference

Deep Neural

Networks

97.3% 2.1%

Real-time detection

abilities

Kumar et al., 2024

Ensemble Learning 95.8% 1.8% Multi-vector attack detection

Wankhade et al.,

2024

SVM 94.2% 2.8% Low com

utational overhea

Saum

a et al., 2024

Random Forest 96.5% 1.5%

Effective for volumetric

attacks

Omarov et al., 2024

3 KEY TAKEAWAYS

The table 3 shows a comparative analysis of various

machine learning models where we can infer that

Tree-based algorithms, particularly Random Forest

and Decision Trees, deliver best performance with

different datasets. Even with default parameters,

these models prove to be highly effective for IoT

Security in Cyber-attack detection with an accuracy

close to 99% in CIC-IoT2023 dataset. (Islam Jony &

Arnib, 2024).

Table 3: Performance comparison of ML algorithms in IoT security.

Algorithm Accurac

Use Case Dataset Reference

Random

Forest

92.72% Binary Classification UNSW-NB15 Saumya et al., 2024

Random

Forest

92.40% Multiclass

Classification

UNSW-NB15 Saumya et al., 2024

Ridge

Classifie

97.00% Real-time Threat

Detection

IoT Network

Traffic

Alomiri&AlShehri,

2024

Decision

Tree

99.19% Cyber-attack

Detection

CIC-IoT2023 Islam Jony&Arnob,

2024

Random

Forest

99.16% Cyber-attack

Detection

CIC-IoT2023 Islam Jony&Arnob,

2024

4 CHALLENGES AND

LIMITATIONS

4.1 Adversarial Attacks

One of the key challenges in using Machine Learning

models is their vulnerability to adversarial attacks.

(Harbi et al., 2024) highlights one of such attacks

named Fast Gradient Signed Method (FGSM) which

is a common attack technique that disrupts and

manipulate the training process by introducing certain

samples, that obfuscate the security systems and

cause the model to make incorrect predictions

bypassing security. (Ibitoye, 2024).

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

412

In order to overcome these attacks, proper

adversarial training is needed. This can be done by

training the model with some adversarial samples

beforehand to ensure robustness against such attacks.

However, this will further increase the training cost

as additional samples need to be used for training

process. This can become a bottleneck as IoT systems

are generally designed with limited resource

availability in mind.

4.2 Dataset Considerations

The UNSW-NB15 dataset has been widely used for

evaluation and validation of proposed systems in real-

world scenarios. Selecting the right dataset with

balanced samples is crucial for validation of your

model. (Omarov et al., 2024). Similarly, CIC-

IoT2023 dataset has proved to be the most reliable

dataset on which accuracy rates of 99.19% and 99.17

percent are achieved by (Islam Jony & Arnob, 2024)

as mentioned in Table 3.

5 FUTURE RESEARCH

DIRECTIONS

With the review of current state of ML in IoT

Security, further researches can be based on resource

optimization in Machine Learning models, aiming to

create light-weight models which can operate

efficiently in constrained environments. Research by

(Wu et al, 2024) have focused on the Pareto-optimal

solutions, which can be the key focus area for future

researchers. Moreover, Hybridization of approaches

can be done to improve performance of existing

systems or to deal with challenges in existing

systems. (Iqbal et al., 2024).

Network-based threats like DDoS needs

innovative detection and mitigation strategies as a

preventive measure in network systems. Future

research can focus on adaptive and intelligent defense

systems which can adapt to varying attack patterns

and can function without manual intervention and

rule base. (Wankhade et al., 2024) have laid the

foundation by introducing these models in IIoT

application, introducing real-time threat response

systems.

According to Harbi et al. (2024), Adversarial

attacks need to be addressed as they disrupt the

fundamental functioning of ML-based security

systems. New systems should balance out the

efficiency and resource usage, for making it easier to

deploy and integrating ML in existing IoT

environments.

Advanced approaches like Federated learning and

collaborative defense mechanisms can be used along

with basic machine learning models, to further

strengthen the system security and focusing on

privacy of data while sharing it in IoT ecosystem.

This will further reduce the risk of data breach and

cyber-attack. Also, Explainable AI in security

solutions can be researched upon as the present focus

is on the model interpretability and the need to make

the model more understandable by the user. (Islam

Jony&Arnob, 2024).

6 CONCLUSIONS

With the comprehensive review of the various

technologies being used to enhance security in IoT

systems, we can conclude by saying that ML has

made a profound impact this domain, particularly in

IDS and threat response systems. Using AI/ML

approaches, we can increase the efficiency of existing

systems and make them autonomous, which will

reduce human dependency and need to manually

program rules for security systems to stay up to date

with latest threats. Table 4 gives the Performance

Comparison of ML Algorithms in IoT Security.

Table 4: Performance comparison of ML algorithms in IoT security.

Trend Key Features Benefits References

Hybrid Learning

Approaches

Combination of Supervised &

Unsupervised Methods

Improved Detection Rates,

Reduced False Positives

Iqbal et al.,

2024

Adaptive ML

Models

Real-time Learning

Capabilities

Dynamic Threat Response Wankhade et

al., 2024

Feature Correlation Optimized Feature Selection Reduced Computational

Overhea

Htwe et al.,

2024

Adversarial Defense Robust Model Training Enhanced Security Against

ML Attacks

Ibitoye, 2024

Evolution of Machine Learning Applications in IoT Security: A Critical Analysis and Future Perspectives

413

The Models will evolve along with the threats.

Recent advancements highlight the exceptional

performance of Random Forest with (92.72%)

accuracy and other improvements like self-learning

algorithms and hybrid systems which have

transformed the scenario of security In IoT. Other

improvements highlight the role of ML in proactive

threat detection, real-time response systems and

threat mitigation.

However, despite of all these benefits, some

challenges like balancing the computational overload,

integration into existing systems, vulnerability

against adversarial attacks etc. is still there. Model

manipulation prevention and feature selection

optimization are crucial for training quality models.

Particularly in case of DDoS attacks, and other

network-based attacks, it is crucial for a ML model to

be adaptive, updated with the latest threats and

resource-efficient. Moreover, hybrid systems are

beneficial to overcome shortcoming and boosting

efficiency of IoT security systems. Resource efficient

defense mechanisms will play a major role in creating

resilient frameworks that are well-suited for various

IoT applications.

REFERENCES

Abdullah, Alomiri., Mohammed, AlShehri. (2024).

Machine Learning-Based Security Mechanism to

Detect and Prevent Cyber-Attack in IoT Networks.

International Journal of Computing and Digital

Systems, doi: 10.12785/ijcds/160148

Akinul, Islam, Jony., Arjun, Kumar, Bose, Arnob. (2024).

Securing the Internet of Things: Evaluating Machine

Learning Algorithms for Detecting IoT Cyberattacks

Using CIC-IoT2023 Dataset. International Journal of

Information Technology and Computer Science, doi:

10.5815/ijitcs.2024.04.04

Atul, Kumar., Kalpna, Guleria., Rahul, Chauhan., Deepak,

Upadhyay. (2024). Enhancing Security in CIC IoT

Networks through Machine Learning Algorithms. doi:

10.1109/iciteics61368.2024.10625536

Bauyrzһan, Omarov., Omirlan, Auelbekov., Bakhytzhan,

Kulambayev., Батырхан, Омаров. (2024). Iot network

intrusion detection using machine learning on unsw-

nb15 dataset. ĶazaķstanBritantehnikalyķ universitetìnì

ņ habaršysy, doi: 10.55452/1998-6688-2024-21-3-48-

Chaw, Su, Htwe., Zin, Myint., Yee, Mon, Thant. (2024).

IoT Security Using Machine Learning Methods with

Features Correlation. Journal of Computing Theories

and Applications, doi: 10.62411/jcta.11179

Haifa, Ali., Vakula, Rani, J. (2024). Machine Learning for

Internet of Things (IoT) Security: A Comprehensive

Survey. International journal of computer networks and

applications, doi: 10.22247/ijcna/2024/40

Kapil, Wankhade., Ravi, Chandra., Kishore, V, Krishnan.,

Srikanth, Arasavilli., Manoj, Chandra, Undi., Amit,

Choudhary. (2024). Mathematical Approach towards

Adaptive Machine Learning Models for Dynamic

Security Threats in Industrial IoT. Advances in

Nonlinear Variational Inequalities, doi:

10.52783/anvi.v27.1359

Olakunle, Ibitoye. (2024). Robust Defenses Against

Adversarial Machine Learning in IoT Security. doi:

10.22215/etd/2024-16076

Xin-Wen, Wu.,Yongtao, Cao., Richard, Dankwa. (2024).

Pareto-Optimal Machine Learning Models for Security

of IoT Applications. doi: 10.1109/smartnets61466.202

4.10577739

Y, M, Saumya., P.V., Vinay., C., Ariel, Pinto., Natasha,

Elizabeth, Correia., Melanie, Crystal, Miranda.,

Joyline, Rencita, Dsouza. (2024). SmartDefend - IoT

Security Using Machine Learning. doi: 10.1109/disco

ver62353.2024.10750744

Yasmine, Harbi., Khedidja, Medani., Chirihane, Gherbi.,

Zibouda, Aliouat., Saad, Harous. (2024). Roadmap of

Adversarial Machine Learning in Internet of Things-

Enabled Security Systems. doi: 10.3390/s24165150

Zafar, Iqbal., Ahthasham, Sajid., Muhammad, Nauman,

Zakki., Adeel, Zafar., Arshad, Mehmood. (2024). Role

of Machine and Deep Learning Algorithms in Secure

Intrusion Detection Systems (IDS) for IoT &

Smart Cities. International Journal of Information

Technology, Research and Applications, doi:

10.59461/ijitra.v3i4.111

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

414