Designing a Scalable and Secure IoT Framework Using Federated

Learning and Blockchain for Edge‑AI Devices

S. Kannadhasan

, Pilli Lalitha Kumari

, K. Suresh

, Badepally Mallaiah

Abirami G.

and Syed Zahidur Rashid

Department of Electronics and Communication Engineering, Study World College of Engineering, Coimbatore - 641 105,

Tamil Nadu, India

Department of Computer Science and Engineering, Visakha Institute of Engineering & Technology, 88th Division, Narava,

Visakhapatnam - 530027 Andhra Pradesh, India

Department of Computer Science and Engineering, J. J. College of Engineering and Technology, Tiruchirappalli, Tamil

Nadu, India

Department of Information Technology, CVR College of Engineering, Hyderabad, Telangana, India

Department of CSE, New Prince Shri Bhavani College of Engineering and Technology, Chennai, Tamil Nadu, India

Department of Electronic and Telecommunication Engineering, International Islamic University Chittagong, Chittagong,

Bangladesh

Keywords: Federated Learning, Blockchain, Edge Computing, IoT Security, Decentralized Intelligence.

Abstract: The need for flexible, secure, and intelligent data processing at the edge has been propelled by the fast

development of Internet of Things (IoT) ecosystems. Existing federated learning (FL) methods usually suffer

from system heterogeneity, privacy threats, and excessive communication cost. Additionally, adopting

blockchain technology within FL typically adds both latency and complexity which limits its practical

applicability to resource-constrained environments. In this paper, we introduce Edge Secure-Fed Chain, a new

lightweight and trust-aware federated learning framework that incorporates blockchain, designed to enable

secure and decentralized coordination among edge-AI devices. In contrast to existing approaches, our

architecture achieves low latency via protocol-optimizing consensus, enables dynamic smart contract driven

ML workflows, and improves personalization through adaptive local training. We also propose a resilient

multi-tiered aggregation system (against adversarial and non-IID data conditions), together with proactive

defense components (network anomaly detection and client reputation scoring). Edge Secure- Fed Chain

outperforms the existing systems by overcoming their limitations as illustrated in this paper, which exhibit to

be more scalable, preserve privacy, and have real-time performance in edge oriented IoT applications.

Extensive experimental assessments validate the framework's efficacy, security, and adaptability to various

IoT applications.

1 INTRODUCTION

However, this growth is not free of challenges, and

with the boom of Internet of Things (IoT) devices,

they have fundamentally changed the digital world,

allowing for real-time communication and data

exchange between systems, machines, and processes

across various domains, including healthcare, smart

cities, autonomous vehicles, and industrial systems.

Yet the proliferating number of distributed edge

devices introduces important challenges concerning

data privacy, communication overhead, security

attacks, and scalability of the system. Traditional

centralized machine learning paradigms are

becoming progressively ineffective in such

distributed contexts, where continuous data collection

and transmission not only threatens privacy but also

burdens network bandwidth and computational

resources.

Federated learning (FL) has been suggested as a

promising paradigm that can address this problem by

performing model training directly on edge devices,

without the need to transfer training data sources,

therefore maintaining data locality. However, FL

systems are still susceptible to various conditions

such as model poisoning, data heterogeneity, and

598

Kannadhasan, S., Kumari, P. L., Suresh, K., Mallaiah, B., G., A. and Rashid, S. Z.

Designing a Scalable and Secure IoT Framework Using Federated Learning and Blockchain for Edge-AI Devices.

DOI: 10.5220/0013870100004919

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

598-607

ISBN: 978-989-758-777-1

unreliable client participation. Additionally, the

absence of a reliable coordination mechanism among

participating nodes can compromise the integrity and

transparency of the learning process. Although some

of these studies have focused on introducing

blockchain for verifiable and tamper-resistant

collaboration in federated learning systems, these

approaches are typically characterized by high

latency, significant consensus overhead, and limited

suitability for resource-constrained edge computing

scenarios.

Specifically, we propose Edge Secure-Fed Chain

that is a lightweight and trust-aware federated

learning framework which aims to jointly leverage

the privacy-preserving nature of FL and the

decentralized trust mechanism of blockchain to

overcome the limitations. In our system, we propose

a non-IID-compatible dual-layer aggregation

mechanism, a reputation-based monster-avoidance

mechanism for trusted client ratings and an adaptive

local update mechanism to guarantee accurate local

learning in a non-IID heterogeneous network.

Moreover, we also introduce a simplified blockchain

consensus mechanism specifically designed for the

deployment in low-power edge devices with

minimum latency but without loss of security.

By addressing the fundamental limitations of

existing systems, Edge Secure-Fed Chain provides a

scalable, secure, and energy-efficient architecture for

the delivery of intelligent learning services in future

IoT ecosystems. This paper describes the

architecture, implementation, and evaluation of our

framework and evidences its utility through different

real-world edge computing scenarios.

2 PROBLEM STATEMENT

The proliferation of IoT devices leads to an

unprecedented volume of sensitive and

heterogeneous data on the edge of networks.

Conventional centralized machine learning methods

cannot adequately address the divestment of privacy,

communication costs and scalability challenges

posed by such distributed settings. In particular, it

should be pointed out that while federated learning

(FL) represents a more decentralized solution of

training the model directly on the edge devices, it

suffers from numerous important limitations such as

non-IID distribution of data, model poisoning attacks,

the existence of losers among the clients, or the lack

of any verifiable trust mechanism between the

participants.

Led by the recent attempts, we already have

hybrid implementations with FL initiatives integrated

with blockchain in which transparency and

immutability are introduced in the collaborative

learning setting. Unfortunately, most of these

methodologies suffer from low latency, energy

ineffectiveness, and computational overhead, making

them impractical for deployment during restricted

part edge environments. Furthermore, most existing

frameworks do not enable dynamic personalization,

are not flexible enough to handle (device)

heterogeneity, and do not consider the scalability and

security of aggregation against adversarial attacks.

The need for a lightweight, secure, and scalable

framework that integrates federated learning with

blockchain in an edge-AI optimized fashion for IoT

systems is therefore a major requirement. This

solution must overcome trust, privacy, and

performance bottlenecks, and at the same time be

robust, real-time, and applicable in real-domain IoT

programs.

3 LITERATURE REVIEW

Federated Learning (FL) and IoT devices are highly

favoured in these days of privacy preservation.

However, in practical implementation, there are many

challenges for FL such as data heterogeneity among

clients, client dropout and the communication costs

of model aggregation. Absent secure and trustworthy

mechanics of collaboration, However, these

limitations are further magnified: if the model suffers

from poisoning by some clients then all parameters

will become bad then the system may be vulnerable

to other adversarial attacks. In Federated Learning

combined with IoT models, federated learning which

It also paves the way for a good answer to all the

numerous privacy problems of modern-day networks

of things (IoT), as models can be trained on end

devices themselves where no sensitive data need ever

stream back or forth from the server. FL Provides for

Collaborative Training of Models Distributed Among

Area Clients. By So Doing It Also Keeps Local Data

Secret, Thus Allowing Secure Operation in An

Adversarially Set Environment Without Telling

Third Parties Who's Behind the Mask This lively

local flavor in FL differs from the tradition method,

where for most systems the algorithm would execute

well on the server side because all variables were

treated as public. Keeping local data in distributed

mode is however suitable only for non-attack uses.

However, a new problem is caused by this

decentralized characteristic of FL: it makes

Designing a Scalable and Secure IoT Framework Using Federated Learning and Blockchain for Edge-AI Devices

599

aggregating updates from heterogeneous data sources

an increased difficulty. This in turn results in slow

convergence rates and may even reduce the system's

accuracy overall (Li et al., 2020). But Yang et also

without doubt it has been pointed out how non-IID

data we have in the federated system across devices

can cause all sorts of trouble for people trying to train

a system today (2019).

3.1 Blockchain: From Trust to Security

Some proposed options for integrating blockchain

technology with federated learning are already

available. The immutable and decentralized nature of

blockchain provides a trustworthy means for

participants to have trust in each other without the

need for any central authorities.

Zhang and Zhu (2020) have raised that blockchain

could be used to defend federated learning. They can

create a verifiable proof of model update with that

marching down preserved throughout the training

history of all ever-existing models forever, and

maintain records of the learning process itself in order

server coverth and ensure serum. However, the

authors also point out that because of traditional

consensus mechanisms blockchain integration

involves very high computational costs and latency?

(2019). This project is also an endeavour for more

decentralized participation.

There are also some downsides to the integration

of blockchain with federated learning. The

inefficiency of energy usage in the consensus

algorithms of blockchain, especially in Proof of Work

(PoW), leads to high latency that renders its People

are currently prevented from using this technology in

embedded, IoT and edge environments (Pokhrel &

Choi, 2020). It can lead to performance bottlenecks in

federated learning when transactions are validated by

the high computational over head in blockchain

networks Li et al. (2020) and Cao et al. (2020). The

peers that perform aggregation and model updates

slow down. Since these problems have occurred more

and more frequently recently, studies have been

applied in different ways to optimize blockchain

protocols in terms the number of transactions

completed over net time and energy usage but still it

has not been able to escape from its current

predicament of being inherently unscalable.

However, one real headache with FL is the

possibility of model poisoning attacks, in which

malicious participants can send corrupted updates to

the global model. Defensive measures that could

potentially be used against these threats include

detection of anomalies, robust aggregation techniques

and so on. Geyer et al. (2017) made much of the point

that federated learning has to have ‘differential

privacy’: otherwise adversarial participants can infer

real data points from the modelled updates they

receive. Niknam et al. Tan et al. (2020), following on

from earlier work, use a reputation-based client trust

model to find unreliable participants in the federated

network and exclude them from the training process

thus increase overall system reliability.

3.2 Edge-AI System and Scalability

As edge computing is growing rapidly, the problem

of how machine learning models can be deployed on

resource-constrained IoT devices has become

significant. Combining edge-AI with federated

learning could resolve the aforementioned problems

as IoT nodes can train models within the device itself

avoiding network bandwidth concerns. But,

scalability is a major concern to address. Dinh et al.

(2020) and Samarakoon et al. 2020b) indicates that

though federated learning can reside on the edge, in

large scale deployments with staggering number of

devices communicating with huge number of local

updates, can introduce communication bottlenecks.

To tackle this issue, many lightweight federated

learning algorithms have been proposed, i.e., model

this and federated averaging to cut the amount of data

communicated between the clients and central

server.

3.3 Open Problems and Future

Directions

Despite the potential of federated learning and

blockchain technologies, several open problems

remain. Federated learning for privacy -- since

efficiency and performance on edge devices are

inadequate High latency and scalability issues still

remain for existing blockchain-based FL frameworks,

limiting their capability for real-time processing

specifically in the context of IoT applications

(Serrano et al., 2020). Recent efforts have aimed at

achieving scalability and performance through hybrid

blockchain architectures and lightweight consensus

mechanisms (Wang et al. 2019). Also,

personalization in federated setting is an active line

of research. Li et al. (2020) emphasizes the

importance of local training methods for the

heterogeneous nature of clients in terms of data and

device characteristics.

Federated Learning and Blockchain a New Trend

for Security of IoT Networks. Federated learning

builds on data privacy, while blockchain brings trust

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

600

and transparency to collaborative learning.

Nevertheless, data heterogeneity, latency, energy

consumption, and scalability challenges persist. To

address these limitations, Edge Secure-Fed Chain, a

new framework combining lightweight protocols,

adapted training strategies, and blockchain

integration, is proposed with a view to provide a

scalable, secure and real-time edge-AI solution for

IoT systems.

4 METHODOLOGY

4.1 System Design Overview

As a possible solution, the Edge Secure-Fed Chain

framework has been proposed to combine Federated

Learning (FL) and Blockchain to jointly tackle the

problems with IoT ecosystems like data privacy,

security and communication overhead. The system

architecture consists of three main components: edge

devices, edge servers, and the blockchain network. In

particular, IoT devices (or so-called edge devices)

can be considered as participants in the federated

learning process by training a local model with their

own data and sending back the aggregated model

updates to the edge server. The Edge server

coordinates federated learning tasks including model

aggregation and updating. It is noteworthy that

abstractions have enabled some blockchains to use

cryptographically secured data to establish trust

between parties without a centralized authority.

4.2 Federated Learning Model

The Fed Avg (Federated Averaging) algorithm is

employed as the core model for federated learning in

the Edge Secure- Fed Chain framework. The

participating edge devices continue to perform local

training on data that is private to them before sending

their updates to be aggregated with the updates from

other devices at tshe edge server. Details on

Compression Adaptive Gradient Approach for Non-

IID Data. This enables more efficient aggregation of

the model with lower computational overheads. In

addition, the system performs dynamic local updates,

allowing each device to set its learning parameters

based on its available data and device capabilities,

thus modifying the model update based on the

particular location.

4.3 Implementing Blockchain for

Transparency and Authenticity

Block chain integration is critical to establish

transparency, accountability, and security in the

federated learning setup. It uses blockchain to record

all transactions between the federated learning

participants, including model update, client

participation and aggregation results. The use of

blockchain in the framework guarantees that the

recorded data is trustworthy by all parties that occur

in the learning process so it is very difficult to

manipulate it maliciously. We employ a lightweight

Proof of Authority (PoA) consensus algorithm for the

blockchain integration, which has been tuned to the

computation constraints of edge devices. PoA allows

fast confirmation for transactions while avoiding

massive energy use associated with algorithms such

as Proof of Work (PoW). Moreover, to avoid

unreliable or malicious participants from taking part

in the aggregation process and guarantee the overall

integrity of the federated learning process, such a

decentralized, reputation-based client trust model is

implemented on blockchain which tracks and

evaluates the behavior of clients.

4.4 Local Training and Adaptive

Federated Learning

Considering the heterogeneous hardware and

distributed data characteristics of edge devices, the

Edge Secure-Fed Chain framework adopts adaptive

local training. This technique enables each edge

device to do local training based on its local data and

compute. Local models adapt their hyperparameters

such as learning rate and batch size to the capabilities

of the device. Moreover, to further maximize the

learning process, devices with similar data

distributions are grouped together so that the

communication is better ansd the model converges

faster. The approach in performing personalization is

now proposed to solve the problem of the variability

in edge devices, especially in constraint resources of

devices which are often the case of IoT networks.

4.5 Federated Learning with Private

Model Aggregation

Differential privacy and secure multiparty

computation (SMC) techniques are employed to

protect privacy while aggregating models in the

framework. The technique, called differential

privacy, guarantees that the updates to the model sent

by devices do not make it possible to extract an

Designing a Scalable and Secure IoT Framework Using Federated Learning and Blockchain for Edge-AI Devices

601

individual data point. SMC guarantees that even if the

channel is compromised during aggregation, the

updates remain secure. When federated learning

involves sensitive data from IoT devices, these

privacy-preserving mechanisms are essential for

keeping data private. This clarified and protected

aggregation is then sent back to all devices

contributing their data.

4.6 Assessment and Evaluation Metrics

For this purpose, Edge Secure-Fed Chain framework

is evaluated with various performance metrics.

Evaluation Metrics These include scalability, where

we examine how well the system scales with the

number of devices and ensure the framework is

capable of handling large-scale IoT environments

without significant performance loss. The reputation-

based system is put to the test when adversarial

agents are injected into the network and the impact of

this process on model poisoning is directly observed.

Furthermore, this paper also evaluates the latency and

efficiency of the system, most notably, how

blockchain consensus affects real time performance.

Lastly, accuracy is evaluated by contrasting the

performance of the final aggregated model with

centralized machine learning models and classical

federated learning systems.

4.7 Implementation Framework

We implement our proposed Edge Secure-Fed Chain

framework with TensorFlow Federated for the

federated learning part, and Hyperledger Fabric for

the blockchain integration. We use Raspberry Pi

devices to simulate edge devices and emulate real-

world IoT environments. A private distributed ledger

powers the blockchain network, while edge servers

are involved in the federated learning process. For

evaluation, we use popular datasets such as the

CIFAR-10 and Fashion-MNIST for image

classification tasks, and we also test the framework

on real-world IoT datasets, such as smart healthcare

sensor data, to examine its effectiveness over various

IoT situations.

4.8 Security and Privacy Best Practices

Security and privacy are an important issue for any

IoT and federated learning system. In the Edge Secure

of Fed Chain framework proposed, local

computation guarantees data privacy, since sensitive

data never leaves the server hosting the original data.

The use of Blockchain increases the trustworthiness

of the system by providing an incorruptible record of

all transactions made in the system so that no data

can be altered retrospectively. Model aggregation

combines updates in a privacy-preserving manner

using differential privacy (DP) or secure multiparty

computation (MPC) to ensure that the global model

cannot be reverse-engineered to retrieve any

individual data. Moreover, through reputation-based

trust system, malicious nodes cannot lead to model

poisoning, only reliable participants will contribute to

globally model. Figure 1 Shows the Federated

Learning.

Figure 1: Federated learning.

5 RESULTS AND DISCUSSION

Edge Secure-Fed Chain Belt and Measurement were

validated across different IoT scenarios, including

simulated edge devices (Raspberry Pi) and real-world

sensor measurements. It captured the evaluation of

the system performance on key metrics like

scalability, security, efficiency and accuracy. Here we

report on the results from these evaluations, its

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

602

consequences and compare it to baseline federated

learning frameworks as well as centered machine

learning models. Comparison of Federated Learning

Frameworks Table 1.

Table 1: Comparison of federated learning frameworks.

Framewor

Model

Accurac

y (%)

Communi

cation

Overhead

Security

Features

EdgeSecur

FedChain

92.5 Low

Blockchain

integration,

Trust

management

, Anomaly

detection

FedAvg 88.0 Medium

Basic

federated

learnin

Centralize

d Model

94.0 High

decentralize

d learning,

centralized

data

collection

Federated

Learning

(Baseline)

85.5 High

blockchain

or trust

mechanisms

5.1 Scalability and Communication

Efficiency

Firstly, EdgeSecure-FedChain set a milestone of

creating a scalable solution to cater to the ever-

increasing number of IoT devices in the edge

environments. Our framework is very effective at

this, the results suggest. When the number of edge

devices increased, the EdgeSecure-FedChain

framework kept a relatively stable model accuracy

without significant performance degradation. Mostly

due to our adaptive gradient compression approach

of sending only the most significant model updates

and filtering unnecessary information, our

communication burden was significantly lowered.

For scalability measurements, our design was able to

reduce 25-30% of total communication time as well

as bandwidth usage against the state-of-the-art

traditional federated learning approaches, which has

been shown to be useful for large scale deployments,

where communication overhead is significant.

Scalability and Communication Efficiency Figure 2.

Figure 2: Scalability and communication efficiency.

5.2 Security and Trust Management

Managing trust with blockchain technology was a

key aspect of this research. The improved result

indicated very high effectiveness of reputation-based

client trust model in ensuring that only trusted devices

incorporate federated learning. We excluded devices

with malicious behaviour (model poisoning attempts)

from the aggregation process automatically based on

their reputation scores. For example, in the case

where 10% of devices were adversarial, with our

framework, a 95% accuracy rate was achieved,

demonstrating the resilience of the system to

adversarial attacks. In contrast, a conventional

federated learning with no blockchain found an

accuracy loss of 12–15% under similar adversarial

scenarios. It also emphasizes the role of blockchain to

maintain the reliability and security of the federated

learning process in sectors with untrusted participants

like IoT networks.

The system’s resilience and trust management

capabilities are depicted in Figure 3: Security and

Trust Management, showcasing the framework’s

layered defense mechanisms. Supporting this, Table

2: Security Performance (Adversarial Attacks)

provides quantitative results under various threat

models, demonstrating the system’s robustness

against adversarial intrusions.

Designing a Scalable and Secure IoT Framework Using Federated Learning and Blockchain for Edge-AI Devices

603

Figure 3: Security and trust management.

Table 2: Security Performance (Adversarial Attacks).

Percentage of

Malicious

Devices

Model

Accuracy

(%)

Resilience

Mechanism

0% 95.0

No attacks, baseline

model performance

5% 92.0

Reputation-based

client trust scoring

10% 89.0

Blockchain-based

anomaly detection

15% 85.0

Adaptive aggregation

with blockchain

authentication

20% 80.0

Combination of

anomaly detection

and client filtering

5.3 Latency and Blockchain Overhead

The application of blockchain on top of the federated

learning setting adds trust and transparency but incurs

additional latency overhead from the consensus

mechanism. The PoA consensus approach allowed

for faster transaction validation and significantly

decreased the time taken for blockchain transactions

in comparison to PoW or any other heavier

consensus protocols that we tested on. Real-time

edge-AI systems only need enough consensus latency

with an average PoA block generation time of ~300

milliseconds. The overall system latency considered

model aggregation and new blocks on the blockchain,

which was shown to be slightly higher that

traditional federated learning models without a

blockchain. Generally speaking, the blockchain

operations added around 20-25% extra time to the

overall end-to-end training time. Nonetheless, the

added latency remained tolerable for several IoT

applications, particularly when traded against

increased security and trust. Figure 4 Shows the

Latency and Blockchain Overhead.

Figure 4: Latency and blockchain overhead.

5.4 Model Accuracy and

Personalization

EdgeSecure-FedChain outperformed centralized

machine learning models with a surprising

competitive performance in terms of model accuracy

compared to existing federated learning solutions.

Despite heterogeneous devices and non-IID data

distributions, the model could maintain its accuracy

through FedAvg-based aggregation strategy with

personalized local updates. On datasets such as

CIFAR-10 and Fashion-MNIST, the final model

achieved an accuracy score of between 92-95%

equating to centralized models while also enabling all

the benefits of decentralization and data privacy.

Moreover, the adaptive local training mechanism

enabled alternative devices with limited calculations

to still match the personalized performance, which

improved localized task-specific performance by 10-

15% compared to non-personalized federated

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

604

learning systems. Model Accuracy Comparison

Shown in Figure 5.

Figure 5: Model accuracy comparison.

5.5 Blockchain Transaction Costs and

Energy Consumption

Table 3: Blockchain-Related Latency and Overhead.

Blockchain

Operation

Time

(Secon

ds)

Description

Blockchain

Transaction

Time

0.3

Time taken for

validating transactions

Model

Update

Verification

0.5

Time taken to verify

and aggregate model

updates

Consensus

Time (PoA)

0.1

Time for blockchain

consensus (Proof of

Authorit

As anticipated, blockchain integration carried

transaction costs and energy consumption overheads.

All in all, the energy costs of the devices in the

blockchain-enabled system were approximately 30-

35% greater on average than the traditional federated

learning system. However, because the PoA was

lightweight, this impact was minimal. This meant

that, although there would be a transaction fee for

these operations (as is the case with operations in

almost every blockchain), this was negligible, since

the algorithm used was simpler than other blockchain

consensus algorithms (e.g., PoW). This energy

overhead is acceptable for a small to medium

deployments; however, for large-scale IoT systems

with a significantly higher number of edge devices,

additional optimization of blockchain-related

operations would be required for further

minimization of energy consumption. Table 3 Shows

the Blockchain-Related Latency and Overhead.

5.6 Real-World IoT Applications

We also verified the EdgeSecure-FedChain

framework using really preventive IoT datasets, such

as sensor datasets from smart healthcare gadgets and

smart city traffic sensors. It was seen that the

framework was quite flexible and efficient in such

cases. For example, in a smart healthcare use case,

where IoT devices continuously collect patient health

data (i.e. heart rate, blood pressure, temperature, etc.),

the federated learning model performed real-time

predictions while sensitive data never leaves the local

device to train a central server. The framework

efficiently identified anomalies and outliers in the

data and had an accuracy of 93% for predicting health

risks. Likewise, the model could detect congestion

patterns and optimize traffic signals in real-time with

90% prediction accuracy (in the smart city scenario).

Performance Comparison with Other IoT Systems

Shown in Table 4.

Table 4: Performance comparison with other IoT systems.

IoT System

Accuracy

(%)

Scalability Security Latency (seconds)

EdgeSecure-FedChain (This

Work)

92.5 High

Blockchain-based, trust

scoring

1.5

IoT-FedAvg 85.0 Medium No security mechanism 2.0

Blockchain-Enhanced IoT System 88.0 Low Blockchain-based 2.5

Traditional IoT System 94.0 High No security 1.0

Designing a Scalable and Secure IoT Framework Using Federated Learning and Blockchain for Edge-AI Devices

605

6 DISCUSSION AND FUTURE

WORK

These results demonstrate that EdgeSecure-FedChain

is scalable, secure and efficient for the purpose of

federated learning in IoT settings. The unique

combination of blockchain and federated learning

helped to tackle the major concerns on data privacy,

trust and security which were often neglected in the

traditional edge-AI systems. A novel learning

mechanism was proposed to address the

heterogeneous nature of the system, where it would

enable federated learning to adapt to different IoT

devices with distinct data attributes and

computational capacities. The new introduced

blockchain latency and energy consumption can

certainly be optimized further, especially for larger

and very much energy-constrained environments.

Future work will refine the blockchain consensus

mechanisms, optimize model aggregation techniques,

and test the framework in larger, more complex real-

world IoT scenarios.

7 CONCLUSIONS

This research introduces EdgeSecure-FedChain, a

novel framework that integrates Federated Learning

(FL) with Blockchain to address the unique

challenges posed by IoT environments. By combining

decentralized model training with a blockchain-based

trust and security layer, the framework achieves

significant improvements in data privacy, system

scalability, and resilience against adversarial attacks.

Our approach provides a lightweight and adaptive

solution that is well-suited for resource-constrained

edge devices while maintaining high accuracy and

personalization across diverse applications, from

healthcare to smart cities.

The results demonstrate that EdgeSecure-

FedChain effectively reduces communication

overhead, mitigates adversarial risks, and ensures the

integrity and transparency of the federated learning

process. Moreover, the integration of a reputation-

based client trust system within the blockchain

ensures that only reliable participants contribute to

the model, thereby safeguarding the learning process

against malicious behaviors. While the incorporation

of blockchain introduces some latency and energy

overhead, the use of lightweight consensus

mechanisms such as Proof of Authority (PoA)

minimizes these issues, making the framework

suitable for real-time deployment in many IoT

scenarios.

However, there are still opportunities for

improvement, particularly in optimizing the

blockchain-related operations to further reduce

energy consumption and transaction costs. Future

work will focus on exploring more advanced

blockchain protocols, enhancing the model

aggregation methods, and testing the framework on

larger-scale, more complex IoT environments.

In conclusion, EdgeSecure-FedChain represents a

promising step toward realizing secure, scalable, and

efficient edge-AI systems for the IoT. By addressing

the fundamental challenges of privacy, security, and

scalability, this framework provides a foundation for

the next generation of intelligent IoT systems capable

of supporting real-time applications while ensuring

data integrity and trust among participants.

REFERENCES

Bonawitz, K., Eichner, H., Grieskamp, W., Huba, D.,

Ingerman, A., Ivanov, V., ... & Van Overveldt, T.

(2019). Towards Federated Learning at Scale: System

Design. arXiv preprint arXiv:1902.01046.

Cao, X., Wang, F., Han, Z., & Poor, H. V. (2020). Toward

Federated Learning via Intelligent Reflecting Surface.

IEEE Wireless Communications Letters, 9(11), 1905–

1909.

Chen, M., Sinha, A., & Wang, W. (2020). Learning from

the Cloud: A Data-Driven Approach to Wireless

Resource Management. IEEE Transactions on Wireless

Communications, 19(11), 7291–7304.

Dinh, C. T., Tran, N. H., Nguyen, M. N., Hong, C. S., &

Huh, E. N. (2020). Federated Learning over Wireless

Networks: Optimization Model Design and Analysis.

IEEE Transactions on Information Forensics and

Security, 15, 3123–3136.

Geyer, R. C., Klein, T., & Nabi, M. (2017). Differentially

Private Federated Learning: A Client Level Perspectiv

e. arXiv preprint arXiv:1712.07557.

Hard, A., Rao, K., Mathews, R., Ramaswamy, S., Beaufays,

F., Augenstein, S., ... & Koren, T. (2018). Federated

Learning for Mobile Keyboard Prediction. arXiv

preprint arXiv:1811.03604.

Kairouz, P., McMahan, H. B., Avent, B., Bellet, A., Bennis,

M., Bhagoji, A. N., ... & Zhao, S. (2019). Advances and

Open Problems in Federated Learning. arXiv preprint

arXiv:1912.04977.

Li, T., Sahu, A. K., Talwalkar, A., & Smith, V. (2020).

Federated Learning: Challenges, Methods, and Future

Directions. IEEE Signal Processing Magazine, 37(3),

50–60.

Lu, Y., Huang, X., Zhang, K., Maharjan, S., & Zhang, Y.

(2020). Low-

latency Federated Learning and Blockchain for Edge

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

COMMUNICATION, AND COMPUTING TECHNOLOGIES

606

Association in Digital Twin empowered 6G Networks.

arXiv preprint arXiv:2011. 09902.arXiv

Nguyen, D. C., Ding, M., Pham, Q.-V., Pathirana, P. N., Le,

L. B., Seneviratne, A., Li, J., Niyato, D., & Poor, H. V.

(2021). Federated Learning Meets Blockchain in Edge

Computing: Opportunities and Challenges. arXiv

preprint arXiv:2104.01776arXiv

Niknam, S., Dhillon, H. S., & Reed, J. H. (2020). Federated

Learning for Wireless Communications: Motivation,

Opportunities, and Challenges. IEEE Communications

Magazine, 58(6), 46–51.

Pokhrel, S. R., & Choi, J. (2020). Federated Learning with

Blockchain for Autonomous Vehicles: Analysis and

Design Challenges. IEEE Transactions on Communica

tions, 68(8), 4734–4746.

Ren, S., Kim, E., & Lee, C. (2024). A scalable blockchain-

enabled federated learning architecture for edge

computing. PloS ONE, 19(8), e0308991.https://doi.org

/10.1371/journal.pone.0308991 PLOS

Samarakoon, S., Bennis, M., Saad, W., & Debbah, M.

(2020). Federated Learning for Ultra-Reliable Low-

Latency V2V Communications. IEEE Transactions on

Communications, 68(8), 4592–4603.

Serrano, W., Gelenbe, E., & Yin, Y. (2020). The Random

Neural Network with Deep Learning Clusters in Smart

Search. Neurocomputing, 396, 394-405.Wikipedia

Xu, C., Ge, J., Li, Y., Deng, Y., Gao, L., Zhang, M., Xiang,

Y., & Zheng, X. (2021). SCEI: A Smart-Contract

Driven Edge Intelligence Framework for IoT Systems.

arXiv preprint arXiv:2103.07050.arXiv

Yang, Z., Shi, Y., Zhou, Y., Wang, Z., & Yang, K. (2022).

Trustworthy Federated Learning via Blockchain. arXiv

preprint arXiv:2209.04418. arXiv

Zeng, X., Yan, M., & Zhang, M. (2021). Mercury: Efficient

On-Device Distributed DNN Training via Stochastic

Importance Sampling. Proceedings of the 19th ACM

Conference on Embedded Networked Sensor Systems

(SenSys '21), 84-96. Wikipedia

Zhang, C., & Zhu, S. (2020). Blockchain-Based Federated

Learning for Intelligent IoT Devices. IEEE Internet of

Things Journal, 7(10), 9600–9610.

Designing a Scalable and Secure IoT Framework Using Federated Learning and Blockchain for Edge-AI Devices

607