IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 26, NO.

12, DECEMBER 2022 5805

Edge Intelligence: Federated Learning-Based

Privacy Protection Framework for Smart
Healthcare Systems
Mahmuda Akter, Graduate Student Member, IEEE, Nour Moustafa , Senior Member, IEEE,
Timothy Lynar , and Imran Razzak

Abstract—Federated learning methods offer secured smart applications for real-time health monitoring, early-stage
monitor services and privacy-preserving paradigms to detection, and cognitive decision-making by gathering patients’
end-users and organisations in the Internet of Things immediate physiological records at any time and from any loca-
networks such as smart healthcare systems. Federated
learning has been coined to safeguard sensitive data, tion via an IoT network. As a result, the demand for IoT systems
and its global aggregation is often based on a centralised is rapidly increasing, with an estimated value of $158 billion by
server. This design is vulnerable to malicious attacks and 2022. Moreover, the recent COVID-19 outbreak has accelerated
could be breached by privacy attacks such as inference and this industry’s hyper-growth [2]. Sensitive information such as
free-riding, leading to inefficient training models. Besides, valuable health insights data collected through high technology
uploaded analysing parameters by patients can reveal
private information and the threat of direct manipulation by devices in IoT systems needs to choose an efficient smart sharing
the central server. To address these issues, we present a platform that meets the high-level privacy of Australian Privacy
three-fold Federated Edge Aggregator, the so-called Edge law with HIPAA [3], [4] compliant.
Intelligence, a federated learning-based privacy protection Currently, federated learning ensures that a user’s data stays
framework for safeguarding Smart Healthcare Systems on their device while applications run particular programs that
at the edge against such privacy attacks. We employ an
iteration-based Conventional Neural Network (CNN) model learn how to handle the data and develop a better and more
and artificial noise functions to balance privacy protection efficient model [5]. Federated learning is an efficient distributed
and model performance. A theoretical convergence bound Machine Learning for privacy preservation for training Machine
of Edge Intelligence on the trained federated learning Learning models. Centralised machine learning in intelligent
model’s loss function is also introduced here. We evaluate
smart healthcare systems may confront several privacy chal-
and compare the proposed framework with the recently
established methods using model performance and privacy lenges, such as stealing users’ data. Because users often lose
budget on popular and recent datasets: MNIST, CIFAR10, control over their data during processing and sharing through
STL10, and COVID19 chest x-ray. Finally, the proposed centralised machine learning algorithms. Alternatively, feder-
framework achieves 90% accuracy and a high privacy ated learning offers a collaborative learning mechanism with
rate demonstrating better performance than the baseline machine learning algorithms by processing data at the user-end
technique.
and sharing the learning model parameters to a central server
Index Terms—Convolutional neural network (CNN), edge instead of raw data.
intelligence, federated learning, Internet of Things (IoT), Cooperation and broadcasting require a central monolithic
privacy-preserving, smart healthcare system (SHS).
intelligent server that executes a model’s changes in a single
unit. Federated learning is a healthcare system that can establish
I. INTRODUCTION an IoT network with smart technology that generates hetero-
geneous data. Attackers in the central aggregator may plan to
HE Internet of Things (IoT) growth enables smart health-
T care systems to step forward for inaccessible patients dur-
ing critical situations like pandemics and ambient assisted living
launch multiple privacy assaults, affecting the entire system at
a single point of failure in its internal components [6]. As a
result, bottleneck traffic experiences strong and well-recognised
(AAL). In this situation, a Smart Health Systems (SHS) [1] offers
performance problems of data flow due to limited bandwidth to
support the volume of data. Besides, all traffic is controlled by the
Manuscript received 8 April 2022; revised 19 June 2022; accepted 15
July 2022. Date of publication 20 July 2022; date of current version 6 De- central aggregator. In terms of privacy, if the global aggregator
cember 2022. This work was supported by the University of New South faces privacy breaches, then all connected users cannot access
Wales (UNSW), Canberra. (Corresponding author: Nour Moustafa.) the application. Privacy attacks will affect the entire system
Mahmuda Akter, Nour Moustafa, and Timothy Lynar are with the
School of Engineering and Information Technology, University of New and other challenges may arise that might cause single point
South Wales, Canberra, ACT 2612, Australia (e-mail: mahmuda.akter@ of failure. Federated learning relies on an integrated entity for
adfa.edu.au; [email protected]; [email protected]). establishing devices, selecting members, and estimating the
Imran Razzak is with the University of New South Wales, Sydney,
NSW 2052, Australia (e-mail: [email protected].). global aggregate in each epoch. These processes constantly bias
Digital Object Identifier 10.1109/JBHI.2022.3192648 the results [7], [8].
2168-2194 © 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: Chandigarh University. Downloaded on April 09,2025 at 04:16:32 UTC from IEEE Xplore. Restrictions apply.
5806 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 26, NO. 12, DECEMBER 2022

Edge Intelligence provides well-organised Artificial Intel- Federated learning overcomes some drawbacks of centralised
ligence placement at edge servers by leveraging large-scale machine learning algorithms by directly processing data on
computation and connectivity capabilities to process data close the client-side and communicating parameters with a central
to the end devices generated [9]. By pre-processing trained server, ensuring data privacy. However, the emergence of new
models before transferring them directly to smart healthcare cyber threats against the federated learning paradigm results in
service providers, Edge Intelligence based on federated learn- various privacy breaches, a significant worry with intelligent
ing supports machine learning in a privacy-preserving manner. embedded systems, particularly the Internet of Things. Because
Edge intelligence can be improved processing to manage over- it takes advantage of the widespread presence of connected
all system privacy collapse for the next generation of smart things in the healthcare system, this mindset has given rise to
healthcare systems. Furthermore, the Federated Edge Intelli- the concept of eHealth, allowing for new healthcare methods
gence (FEI) architecture [10] enabled machine learning-based and solutions [19].
smart services and applications to make federated learning more Furthermore, the attackers’ purpose is to compromise users’
energy-efficient. confidentiality, availability, and integrity [20]. The insertion of
Although edge intelligence-based federated learning has vari- carefully prepared hostile instances into traditional Machine
ous advantages over machine learning-based intelligent applica- Learning models made possible by acquiring training data from
tions in IoT systems, establishing such machine learning-driven the public domain renders traditional Machine Learning models
intelligent applications with a high accuracy score necessitates vulnerable to attackers. It allows the model to readily adjust
greater data privacy protection. Datasets exposed to third-party to the attacker’s desire [21]. In addition, the central aggrega-
edge servers may result in significant privacy infringement [11]. tor might sometimes be honest but curious and wish to use
Furthermore, current privacy standards and regulations limit parameterised models for analysis or give access to an unautho-
access to sensitive data in intelligent applications [12]. As a rised organisation. The curator might assess the client’s nature
result, offering privacy-preserving access to large databases and track which shared a specific updated model. Suppose all
while assuring increased data analytics is crucial. Most current connected clients’ data privacy will face a significant threat
research focuses on maximising IoT communication resource if privacy attacks violate the central aggregator. A massive
allocation and distributing computational burden across edge amount of IoT data model aggregation and broadcasting from
computing [13]. As a result, incorporating edge intelligence into a central server might face significant traffic congestion that
the edge computing layer could help to improve overall resource will impact overall system performances, including energy time
optimisation. and accuracy.
Anonymisation, encryption, and randomisation methods have According to recent studies, federated learning as distributed
been studied to produce privacy-preserving strategies [14]. They deep learning (DDL) can provide better privacy protection than
are, however, ineffective at preventing the leakage of sensitive centralised deep learning because it only trains on shareable pa-
data from Smart Health Systems in an IoT network. Differential rameters [22]. The resource consumption of a federated learning-
privacy could take advantage of manufactured noise in this situ- based algorithm is closely related to several key parameters,
ation, causing each client’s information to be perturbed locally including the number of edge servers participating in each round
and only sharing a randomised version with the aggregator. A of global model coordination, the number of local computational
Gaussian noise-adding methodology is more efficient for data steps performed by each edge server, and the total number of
privacy in shared models in federated learning-based privacy- global coordination rounds required to achieve a given model’s
preserving strategies. It will protect users’ data from poisoning, accuracy [23]. Although studying the effects of these parameters
backdoor attacks, and other threats. There have been some has been done, most of these studies have concentrated on
ways to handle privacy preservation for large data sources. A optimising a single parameter that could only affect data pro-
central aggregator in federated learning is in charge of gathering, cessing and computation at edge servers. Furthermore, because
aggregating, and broadcasting a model to many clients [15]. the hyper-parameterised model contains original data, a client’s
On the other hand, a breached central server could com- data need must be disrupted with additional noise to protect
promise privacy through direct link capability or client iden- sensitive data. Traditional perturbation techniques, on the other
tifiability [16]. IoT devices are also not limited to a specific hand, have some limitations.
number in real-world data generators. Controlling a model’s Moreover, Edge computing enhances the performance of con-
massive inundation to its central aggregator also produces data nectivity, speed, and reduces latency of data transfer [24] from
traffic and congestion. As a result, loading the distribution and user to server. To manage the privacy preservation in edge servers
pre-processing of the model is required to improve the middle using machine learning techniques is necessary; the realms of
association’s privacy preservation. diagnosis, prognosis, and spread control, as well as assistive
Motivational Scenario—Federated learning has been ap- systems, monitoring, and logistics, have been split into various
plied to IoT devices with high-functioning designs to reduce IoT-based healthcare applications [25]. The characteristics of
communication rounds during model training. Moreover, it out- smartphones to control and monitor implantable medical de-
performs classic distributed learning approaches for IID (Inde- vices, as well as some practical security methods that improve
pendent and Identically Distributed) and non-IID data distri- the security and privacy of patients, is addressed in [26].
bution methods [17]. Non-IID training data necessitates fewer Key Contributions - This study proposes a novel conceptual
data transmissions from IoT devices to maintain privacy [18]. three-fold hierarchical privacy-preserving framework known as

Authorized licensed use limited to: Chandigarh University. Downloaded on April 09,2025 at 04:16:32 UTC from IEEE Xplore. Restrictions apply.
AKTER et al.: EDGE INTELLIGENCE: FEDERATED LEARNING-BASED PRIVACY PROTECTION FRAMEWORK FOR SMART 5807

a federated edge aggregator over a typical federated learning

method for IoT-based Smart Health Systems. By instituting edge
aggregator between client and server to process learned model
before global updating as Edge Intelligence. We assume that
IoT-based smart healthcare systems obtained heterogeneous data
from various data sources such as rural hospitals, remote pathol-
ogy, and home-based self-isolated patients’ health monitoring
data. To ensure data variation for privacy-preserving evaluation,
we experiment on the non-IID data distribution of the dataset.
By leveraging the proposed Federated Edge Aggregator method,
individual health IoT data is learned in its smart device using the
Convolutional Neural Network (CNN) model. After perturbing
its parameters with differential privacy, they are sent to the edge Fig. 1. A Complete scenario of the Federated Learning procedure.
layer.
After several iterations, until the model converged, this
blended model was sent for Global aggregation to the central Finally, Section VI accomplishes the work with the conclusion
healthcare service provider. To classify smart health monitoring and delivers future research directions.
groups, a couple of edge aggregator is distributed where each
edge aggregation will process each sector, such as pathology
or isolated patients’ vitals data. The central healthcare provider II. BACKGROUND AND RELATED WORKS
aggregates all the received models for final update and broadcast. A. Federated Learning
To our best knowledge, Edge Intelligence for privacy preser-
vation in federated learning and differential privacy has not Since 2015, the concept of federated learning has attracted
been investigated before. In our context patients, monitoring and much attention [27]. It holds much promise in healthcare data
data learning remains in patients’ end smart devices and share analytics [28]. Because data owners in a federated learning
data to the service provider for early detection before critical system are not required to submit their data to the central aggre-
situation without traceability of the client’s identity and avoiding gator, learning is done through a sharing mechanism rather than
other privacy attacks. This paper also shows the robustness of raw data transmission, ensuring confidentiality and data privacy
the proposed method by evaluating existing popular datasets in at a cost significantly greater than the loss of accuracy [15].
both Grayscale and RGB compared with the baseline traditional Each node and central aggregator train a combined Neural
federated learning method. The contributions stated below are Network(NN) model in federated learning. To ensure that the
the focus of this paper: neural network model converges, the nodes communicate their
—We propose a hierarchical three-fold Federated Edge Ag- gradients to the central server, aggregating all gradients and
gregator (FEA) framework with differential privacy applied with providing updates to each node [29]. Ti (i = 1, 2, . . ., n) have
edge intelligence. agreed on a model design if each branch has n IoT nodes, as
—We design a framework that ensures data privacy by lever- shown in Fig. 1.
aging the Gaussian noise adding mechanism before passing The mathematical concept of federated learning is described
model parameters under certain noise perturbation levels. as follows. The function f (x, N ) is a Neural Network model,
—We show the federated learning methods privacy attacks where x is the inputs from the sensors and N the model’s
classifications and possible defence against those threats with parameters, which contains all the biases and connections among
extensions. all the neurons. If the nodes in the branch Ti hold the training set
—We establish an optimal privacy-preserving method to bal- (Di ={(xi , yi ) | j= 1,2,...,P }, where xi is the attributes, yi the
ance privacy budget, model accuracy, and training, test loss class labels and P the size of Di ), the loss function Lf (Di , N )
calculation in different data distribution strategies, and visualise can be defined as:
the trade-off between privacy protection level and convergence
1 
performance with different datasets and compare with the recent Lf (Di , N ) = (yj − f (xj , N ))2 (1)
P
method. (xj ,yj )
—We showed that our proposed method accounts for higher
performance than the baseline method and different hyperpa- The aim of training the NN model is to obtain the gradient to
rameter settings. update N to reduce the amount of the loss function Lf (Di , N ).
The paper is organised as follows. Section II focuses on The stochastic gradient descent (SGD) is used to calculate the
background and related work based on recent research to the participant Ti’s gradient (wi ) as:
proposed study. The essential perceptions and schemes used in
the proposed method are described in Section III. Section IV wi = ∇Lf (D∗ i, N ) (2)
explains the proposed Federated Edge Aggregator method and
mathematical analysis. Section V demonstrates the experimental where ∇Lf is a derivative of the loss function (Lf ) and D∗ i a
results and performance evaluation of the proposed methods. random subset of Di . Then, participants upload their gradients

Authorized licensed use limited to: Chandigarh University. Downloaded on April 09,2025 at 04:16:32 UTC from IEEE Xplore. Restrictions apply.
5808 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 26, NO. 12, DECEMBER 2022

TABLE I solution for creating smart health systems [33]. Many disease
SUMMARY OF KEY NOTATIONS
detection schemes can be deployed using smart sensors and
connecting with smart devices in the user’s home environment.
These schemes collect users’ health-related data and process
it in the data analysis cloud. As a remote detection system,
smart health systems can greatly aid isolated patients who have
problems moving or are incapacitated. Smart healthcare has seen
several success stories because of the amount of user-health data
accessed by IoT devices and recent breakthroughs in Machine
Learning [34].

C. Differential Privacy
Because of its outstanding privacy guarantees, Differential
Privacy is commonly used. The privacy leakage of single data
belonging to a person in a dataset is addressed when some
information from that dataset is made publicly available. So
that to establish a strong privacy assurance, differential privacy
brings a robust statistical method. As discussed below, differ-
ential privacy is defined as a mathematical computation and
algorithmic.
Definition 1: A randomised function F gives -differential
privacy if for all data sets d1 and d2 differing on at most one
element, and all C ⊆ Range(F),
to the central server for aggregation as:
n
 P r[F (d1) ∈ C] ≤ exp() × P r[F (d2) ∈ C] (4)
w= wi (3)
Where Pr[F(d1)] is a probability of randomised function F of
i=1
dataset d1 and Pr[F(d2)] is a probability of randomised function
Finally, the central server sends the aggregated updates to F of dataset d2. F is a mechanism that meets this description,
every member, and members update model parameter N = which focuses on worries about personal information leaking
N-η · w/n after receiving w, where is the learning rate. If the even if the user removes data from the data collection. No
closure conditions are not encountered, continue the next round outputs are more or less likely. C is all subset of F and is a
of federated learning. Because this approach does not rely on constant specified by the user, and exp is the base of the natural
the exchange of client data – only the updated model parameters logarithms.  is a measure of The IoT is a massive, embedded
– it provides immediate client privacy. Furthermore, Federated network of connected physical objects and people. IoT devices
Aggregator algorithms such as FedAvg and FedProx work ef- usually generate data and exchange data through a network.
fectively in the presence of non-iid client data distributions [30]. Nowadays, IoT is emerging with a tremendous privacy budget,
Table I describes the mathematical symbols briefly. where less value indicates less privacy loss. If  is 0, that means
no privacy loss.
B. IoT-Enabled Smart Healthcare Systems Definition 2: A randomised function F gives (, δ)-differential
The IoT is a massive, embedded network of connected privacy if for all data sets d1 and d2 differing on at most one
physical objects and people. IoT devices usually generate element, and all C ⊆ Range(F),
data and exchange data through a network. IoT is emerging P r[F (d1) ∈ C] ≤ exp() × P r[F (d2) ∈ C] + δ (5)
with tremendous revolutionisation in different applications,
for example, smart cities, smart medical systems, Industry (, δ)-differential privacy says that for every pair of neighbour-
4.0, agriculture 4.0, etc. For standardisation of IoT on these ing databases d1, d2, it is improbable that the observed value
applications, Artificial Intelligence brings IoT in unique F(d1) will be much more or much less likely to be generated
deployment, smart management, smart prediction, and decision when the database is d1 than when the database is d2. It ensures
making and introduces advanced technology that requires that for all adjacent d1, d2, the absolute value of the privacy loss
privacy involvement. Edge nodes are essential as brokers to will be bounded by  with a probability of at least 1 − δ.
mitigate privacy issues and efficiently manage massive IoT
data integration [31]. Edge intelligence improves and speeds III. PRELIMINARIES
performance while safeguarding user privacy [32].
A vast population means more chronic diseases and criti- A. Edge Intelligence
cal situations requiring frequent visits to healthcare providers, This section briefly discusses the underlying principles of
higher medical treatment expenses, and the need to monitor vital Edge Intelligence and the artificial noise-adding processes
status. As a result, IoT technology appears to be a potential utilised to construct the suggested work. Edge Intelligence is

Authorized licensed use limited to: Chandigarh University. Downloaded on April 09,2025 at 04:16:32 UTC from IEEE Xplore. Restrictions apply.
AKTER et al.: EDGE INTELLIGENCE: FEDERATED LEARNING-BASED PRIVACY PROTECTION FRAMEWORK FOR SMART 5809

the concept of deploying computationally expensive intelligent

models (primarily machine learning, deep learning, and data
analytics) on edge clouds rather than the central cloud [36]. Edge
computing may deliver enormous quantities of data directly
to the cloud. It can receive data from the edge layer before
it goes to the aggregator and determines its importance [37].
However, edge computing is a type of computing between smart
end-devices at the network’s edge and traditional cloud or data
centres [38]. Where Fog computing allows relevant data storage,
irrelevant data deleting and analysing data for remote access, and
localised learning models. Instead of depending on a centralised
server, an edge layer has been proposed to pre-process massive
Fig. 2. Classification of privacy preservation techniques.
IoT data before uploading and storing resources closer to the
devices where data is gathered.
Edge Intelligence [39] significantly reduces communication Hence, we consider our proposed method’s Gaussian mecha-
costs and simultaneously plays data consumer and data gen- nism for adding noise. In the Gaussian mechanism, Additive
erator by executing multitasking such as caching, managing, noise is combined with numeric queries and both IID and non-
offloading, and service delivery. For efficient responsibilities IID entries in the Gaussian technique. When the wireless fed-
in the network, edge nodes must be well-designed to meet erated learning method is employed for its iterations, it may be
major service requirements, such as dependability, security, and used for Local Differential Privacy to acquire the entire privacy
privacy protection. Due to low response times, communication leakage [44]. Gaussian mechanism returns several functions
costs, and higher privacy traits, edge intelligence is a great f(d1), and the definition F(d1) implies with (, δ) -differential
emerging advantage for clients to process sensitive data before privacy.
uploading it to a centralised aggregator server.
F (d1) = f (d1) + ν(σ)2 (6)
Most of the research has focused on determining the privacy
B. Threat Model
loss for a given noise distribution and the composition of privacy
Unlike traditional machine learning systems, federated learn- losses. For the Gaussian noise, here used noise scale that the
ing systems must contend with three potential adversaries: moment accountant maintains track of a bound on the random
clients, aggregators, and outsiders or eavesdroppers. In general, variable’s moments of privacy loss, ensuring efficient noise
numerous protection measures defend federated learning from distribution.
a wide range of threats, reducing the chance of privacy issues. E. Moment’s accountant theorem Nowadays, the study of
privacy loss for a specific noise distribution and the composition
of privacy losses has gotten much attention. We prefer noise scale
C. Privacy-Preserving
σ for the Gaussian noise with sampling probability p, the ratio
Privacy preservation is an essential demand for users, clients, of M samples and N numbers.
and service providers. Privacy might need any data that can M
be sourced from the Internet of Things, social networks, In- p= (7)
N
telligent business applications, databases, documents, the web,
etc. [41]. In cyber security, privacy might include personal After T number of steps, privacy loss  is accounted for in
information, individual privacy, interactive privacy, and com- differential privacy with constant c,
munication privacy [42]. To protect digital assets’ privacy, cy-  < cp2 T (8)
ber security mechanisms are defined as Privacy preservation.
Different types of privacy preservation approaches are grouped For (, δ)-differentially private stochastic gradient descent
into five categories [14] encryption-based, perturbation-based, (SGD) algorithm, we choose,
  
authentication-based, differential privacy (DP), and blockchain-
c p T log( 1δ )
based. A wide range of research is going on to provide privacy σ> (9)
preservation schemes. There have been numerous studies related 
to privacy preservation in smart healthcare systems. We can We consider the moment accountant theorem in our method.
classify as shown in Fig. 2.
IV. FL-BASED PRIVACY PROTECTION METHOD
D. Noise Adding Before Aggregation A. System Overview
There are four popular noise mechanisms in differential pri- By integrating edge aggregator at the intermediate level with
vacy: Exponential, Laplacian, Gaussian, and Hybrid mecha- traditional FL architecture, the proposed 3-fold Federated Edge
nisms [43]. The Gaussian mechanism mostly complies with Aggregator framework ensures individual client privacy and data
(, δ)-differential privacy rather than pure -differential privacy. privacy, as shown in Fig. 3. This framework illustrates a typical

Authorized licensed use limited to: Chandigarh University. Downloaded on April 09,2025 at 04:16:32 UTC from IEEE Xplore. Restrictions apply.
5810 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 26, NO. 12, DECEMBER 2022

Fig. 3. A Conventional scenario to train and deploy federated learning

in Smart Healthcare Systems.

Fig. 5. Flowchart of the proposed 3-folds FEA method.

r Step 1: Global aggregator initialises edge aggregators,

hyperparameters, model parameters, noise parameters,
dataset, and distribution method.
r Step 2: Edge aggregators initialise corresponding IoT
clients, passing model parameters and assigned data.
r Step 3: IoT clients load the training set and label to train
the model, calculate training loss, employ optimisation
technique, load model parameters, add Gaussian noise to
Fig. 4. Proposed framework of three-fold edge aggregator-based fed- each model parameter and update the model
erated learning in smart healthcare systems. r Step 4: Edge aggregator accumulates model parameters
of the corresponding set of IoT clients. Edge aggregator
scenario for training and deploying federated learning using broadcasts updated model to the connected IoT clients
Federated Edge Aggregator on an IoT-based Smart Healthcare edge aggregator calculates testing accuracy test data and
System, collaborating with application sectors, an inquiry anal- corresponding label.
yser, and adversaries. Individual edge aggregator is responsible r Step 5: Repeat Step 2-Step 5 until the model con-
for aggregating and broadcasting the respective Client’s Model verged/stabilised.
for some iterations to produce a better-converged model at the r Step 6: Global aggregator accumulates edge aggregators
edge. After aggregating models from individual clients, the edge model parameters. The global aggregator broadcasts an
aggregator will pass the model to the Global Aggregator (GA). updated model to the connected edge aggregators that
The proposed method in Fig. 4 visualises the steps that are calculate overall testing accuracy from test data and cor-
indicated by circled numbering, and labels are titled: Top Level, responding labels.
Intermediate Level, and IoT Level. The complete functionality Edge intelligence is a hybrid of edge computing and artifi-
of the levels is described below: cial intelligence that takes advantage of end devices’ storage,
r Top Level: The top-level contains one GA as a tree hierar- networking, and processing capabilities to enable edge caching,
chy of the proposed 3-folds Federated Edge Aggregator. model training, and inference closer to the source of data. So,
The primary responsibility of GA is to execute step-1 and instituting edge intelligence at the intermediate level brings a
step-7. viable solution to address privacy attacks. Edge intelligence
r Intermediate Level: The proposed 3-folds FEA introduces allows data owners to process models without transmitting their
a set of EAs at the intermediate level to preserve IoT sensitive parameters directly to third-party servers.
clients’ privacy and their data through steps 2–5. The flowchart in Fig. 5 represents the overall workflow of
r IoT Level: The bottom level of the proposed 3-fold Feder- the proposed technique, which starts with the global aggrega-
ated Edge Aggregator contains a set of clients, including tor, edge aggregator, and client initialising. And the process
laptops, TAB, SmartPhone, and Sensing Devices. The shows for the N epoch until the model converged. And the
primary function in this level is to execute step 3 and conception of differential privacy, in which each client per-
communicate with the Intermediate level. turbs its learned parameters locally by purposefully adding
The steps-by step descriptions of the proposed 3-folds FEA noise before uploading them to the server for aggregation, is
framework are as below: used to avoid information leaking successfully. Fig. 5 shows

Authorized licensed use limited to: Chandigarh University. Downloaded on April 09,2025 at 04:16:32 UTC from IEEE Xplore. Restrictions apply.
AKTER et al.: EDGE INTELLIGENCE: FEDERATED LEARNING-BASED PRIVACY PROTECTION FRAMEWORK FOR SMART 5811

method randomly as-signs the training data. The non-IID-based

method ensures a different dataset for each client, whereas the
IID method may assign a similar dataset. It is to be noted that the
the whole working procedure of the method by evaluating an proposed method ensures robustness by employing the non-IID
existing dataset. method along with the IID method in terms of data distribution.

B. Proposed Three-Fold FEA Algorithm C. System Analysis

The pseudocode of the proposed three-fold Federated Edge This section investigates the computational complexity and
Aggregator is presented in Algorithm 1. This algorithm the level of privacy of IoT data of the proposed 3- fold Federated
presents a complete procedure from IoT clients’ assignment Edge Aggregator model.
to the final aggregated model. Firstly, a set of EA (Ek ← Analysis of Computational Complexity: To analyse the com-
{E1 , E2 , . . ., En }) and corresponding Clients ((Ck , m ← putational complexity of our system employing Big notation. At
{C1,1 , C1,2 , . . ., C1,m }, . . ., {Cn,1 , Cn,2 , . . ., Cn,m })) is ini- first, we investigate the computational complexity of the baseline
tialised with default hyperparameters and dataset. Then based on method that does not have an intermediate layer. We use a basic
the data distribution method (IID or non-IID), data is distributed federated learning approach as a baseline method and add noise
among a certain number of clients corresponding to each edge before aggregating it to the central server. The complexity of
aggregator as described in Algorithm 2. federated learning is (I *N *G), where I represent the number of
For each iteration j(j = 1, 2, . . ., N ), individual IoT Clients iterations, N is the total number of clients, and G is the centre
will load their assigned data, train the model, compute training number.
loss, apply optimisation to reduce training loss, and add Gaus- Secondly, we analysed the time complexity of this Baseline
sian noise to the model parameters. Once all the clients execute method as (C), where C refers to the number of classes. In
the procedures, the corresponding edge aggregator will aggre- Our proposed method, the computational complexity involves
gate clients model, broadcast updated model among respective the IoT, Intermediate, and Top layers. The complexity of all
clients and compute testing accuracy for a certain number of layers is (I+E+N/n+G). Here, I mean the number of Iteration, E
iterations, N. After that, the global aggregator will aggregate means edge aggregator, N/n implies N number edge aggregator
EAs Model, broadcast the updated model to the corresponding distributes n clients, and G is a global aggregator. Additionally,
EAs and finally, compute the global model accuracy. This joint the worst-case scenario is estimated by (A). Similarly, the time
resource allocation and incentive design framework for edge complexity of our proposed method is (n), where n is the number
intelligence reduce reliance on a central controller. of EA used for enhancing privacy preservation.
Based on the data distribution method, Algorithm 2 distributes Analysis of the suitability of resource-constrained devices:
the amount of training data with corresponding labels to each Machine learning methods are challenging to implement
client. The non-IID method firstly sorts the training dataset as because they demand sufficient CPU power and big storage
per labels and then distributes it among Clients; however, the IID memory. To investigate the suitability of resource-constrained

Authorized licensed use limited to: Chandigarh University. Downloaded on April 09,2025 at 04:16:32 UTC from IEEE Xplore. Restrictions apply.
5812 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 26, NO. 12, DECEMBER 2022

devices, our method supports friendly, lightweight end devices TABLE II

HYPERPARAMETERS WITH VALUES USED FOR OUR FEDERATED EDGE
for end data processing. Additionally, the edge intelligence AGGREGATOR METHOD ANALYSIS
device at the edge server and central service provider is used
for overall model updating and analysis.
Analysis of the attack resistance via edge intelligence: the
essential part of privacy preservation is investigated here. Ini-
tially, we assume that our edge aggregator is trusted and located
near edge nodes. A couple of edge aggregator is distributed
for purpose-based (rural hospital or pathological laboratory)
clients’ IoT network. All the edge aggregator is connected
with the global aggregator (central telehealth service provider).
Here we consider global aggregator might be honest but has
curious behaviour. Clients directly communicate with a global
aggregator in a traditional federated learning mechanism.
To show the performance of medical images to justify
In this case, an adversary might trace or localise the client’s
the smart healthcare system, we choose the newly released
identity by attacking a global aggregator. Similarly, a global
COVID19 Chest X-RAY image dataset with 4 labels 20685
aggregator might be able to analyse clients’ patterns because it
training set (COVID19: 3496, Lung Opacity: 5892, Normal:
communicates directly to get model parameters for update every
10072, Pneumonia:1225) and 240 testing set, Whereas Input
time. At the same time, if the global aggregator faces privacy
image dimension 299 ×299. Noted that, due to our machine’s
breaches, all the related client’s privacy might fall apart. Hence,
memory limitation, we had to reduce image dimensions during
edge intelligence performs a broker for model pre-processing
the experiment and resized image dimension was 32×32, which
at the edge layer before sending it to a global aggregator. The
might impact the overall accuracy calculation in our method
edge aggregator sends a blended model after running a certain
evaluation.
iteration of several clients’ updating models. The global aggre-
gator will be unaware of the client’s information. Moreover, the
whole system collapse probability becomes reduced because all C. Data Distribution
the clients are not connected at a single point. The following two data distribution processes are investigated
for existing datasets to assess the method’s robustness. These are
V. EXPERIMENTAL RESULTS AND PERFORMANCE explained below:
EVALUATION —IID: The term “independent and identically distributed”
(IID) refers to a sample of independent events in which all
A. Environment Setup
items are taken from the same probability distribution. In this
We perform our experiments using Microsoft Windows 10 distribution, the data is handled and shuffled in random order.
Home with Intel CoreTM i7 processor (2.6–5.0 GHz), 16 GB So that two clients can get similar class label data. In IID
RAM NVIDIA GeForce GTX 1650 Ti with Max-Q design distribution, divide the training data into equal-sized shards and
graphics processor HP Laptop, and 1 TB SSD. Python 3.6, allocate one shard to each client [17].
Pysift, Pytorch library, and PyCharm 2021.1.2 are used as de- —Non-IID: Heterogeneity and hierarchical coupling connec-
velopment tools. tions among the data and clients are involved in non-independent
and identically distributed (non-IID) data. The data must first be
B. Datasets sorted before being provided to all clients so that comparable
class labels are not assigned to them [45]. Non-IID data distri-
Federated learning models with Differential Privacy are im- bution is widely employed when user devices generate data.
plemented and tested on a dataset from the Modified National Most of the experimental work on non-IID datasets has fo-
Institute of Standards and Technology (MNIST) database [13]. cused on label distribution bias, which consists in partitioning a
The testing accuracy and computational durations levels are flat with an existing dataset based on labels to create a non-IID
investigated as a baseline technique by adjusting different hy- dataset. In non-IID distribution, the unbalanced local data sam-
perparameters, as indicated in Table II. ples, labels, and features as the specific probability distribution
The MNIST database is enormous, with handwritten digits of training examples stored at local nodes can substantially im-
extensively used in machine learning training and testing. It pact the training process’s performance. In non-IID distribution,
includes 60,000 training and 10,000 testing grayscale images data divides are nonoverlapping and balanced, resulting in the
with a 28 × 28 × 1 resolution and ten output classes labeled 0 same number of data points for each client [17].
to 9. In the RGB dataset, we use the popular CIFAR10 dataset
containing a 50000 training set and 10000 testing set with ten
labels where the input image dimension is 32 × 32. The STL10 D. Model
dataset contains 5000 training sets and 8000 testing sets with 10 The study was carried out using a two-layer sequential Con-
labels, where the input image dimension is 96 x 96. Due to the volutional Neural Network (CNN) model with a fully connected
low amount of training set of STL10 dataset, overall learning layer on top. Each layer has a convolution with a kernel size of
may impact. 5X5, batch normalisation, and the activation function RELU.
Authorized licensed use limited to: Chandigarh University. Downloaded on April 09,2025 at 04:16:32 UTC from IEEE Xplore. Restrictions apply.
AKTER et al.: EDGE INTELLIGENCE: FEDERATED LEARNING-BASED PRIVACY PROTECTION FRAMEWORK FOR SMART 5813

Fig. 6. A two-layer sequential Convolutional Neural Network (CNN)

model with a fully connected layer.

Fig. 8. Training error comparison for IID data distribution of the MNIST
dataset: Proposed vs. Baseline.

Fig. 7. Training error comparison for non-IID data distribution of the

MNIST dataset: Proposed vs. Baseline.

Fig. 9. Testing accuracy comparison for non-IID data distribution of the

Fig. 6 shows the procedure of the CNN model, which is how MNIST dataset: Proposed vs. Baseline.
patients’ data will be learned from their own devices.
As a feed-forward neural network convolutional neural net-
work(CNN) use filters and pooling layers, whereas some other
Deep learning model feed results back into the network. Ba-
sically, the input size and the resulting output size are fixed
in CNN. It automatically identifies the critical features without
human supervision with limited data whereas the Deep learning
model requires huge data to perform better. And deep learning
is expensive and has training complexity.
The non-IID method was implemented to split and assign
training data to each client. It is to be noted that the non-IID-
based technique firstly sorts training data as per label and then
gives it to clients; hence, it is ensured that each client will
get a maximum of three-digit variations data. The IID-based Fig. 10. Testing accuracy comparison for IID data distribution of the
technique, on the other hand, does not require sorting. MNIST dataset: Proposed vs. Baseline.

E. Experimental Results and Discussion r Testing Accuracy Analysis: Fig. 9 and Fig. 10 illustrate
We performed several experiments with different datasets to the proposed and baseline methods’ testing accuracy in
study the impact of differential privacy on model accuracy under terms of the MNIST dataset with two distribution methods:
various privacy loss settings. non-IID and IID. Fig. 9 shows the reported 80% and
r Training Error Analysis: Fig. 7 and Fig. 8 demonstrates 90% testing accuracy at the 20th and 75th communication
the proposed and baseline methods’ training error in terms rounds, respectively, for the proposed method; however,
of Mean Absolute Error (MAE). We consider the MNIST 62% and 78% are reported for the same communication
dataset with two distribution methods: IID and non-IID. rounds of the baseline method in terms of non-IID distri-
The proposed method reported lower training error than bution.
the baseline method for certain communication rounds in Fig. 10 shows the reported 80% and 90% testing accuracy
both data distribution methods. at the 17th and 61st communication rounds, respectively,

Authorized licensed use limited to: Chandigarh University. Downloaded on April 09,2025 at 04:16:32 UTC from IEEE Xplore. Restrictions apply.
5814 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 26, NO. 12, DECEMBER 2022

Fig. 12. Performance analysis of our proposed method with different

datasets (MNIST, CIFAR10, STL10, and COVID19 chest x-ray).

Fig. 11. Box Plot of Hyperparameters different settings in terms of

testing accuracy and training loss dem onstration for Baseline and Pro-
posed method. (a) Iteration Settings. (b) Epsilon Setting. (c) Learning
Rate Setting.

for the proposed method; however, 61% and 76% are re- Fig. 13. Levels of testing accuracy vs of proposed and baseline meth-
ported for the same communication rounds of the baseline ods for different data distributions: non-IID based with (a) 5 IoT clients,
method in terms of IID distribution. As can be seen, in (b) 10 IoT clients, and (c) 20 IoT clients; and IID based with (d).
both cases, the proposed method reported higher testing
accuracy than the baseline method for the entire commu-
nication rounds.
r Performance Analysis based on Hyperparameter Dif-
ferent Settings: The box plots in Fig. 11 demonstrates
the proposed and baseline methods’ testing accuracy and
training error in terms of the hyperparameter’s different
settings. We consider the MNIST dataset in this regard
for the performance evaluation. The proposed method
reported a higher median value (bold red hyphen marked
on the graphs) than the baseline method for all the set-
tings of iteration (Fig. 11(a)) and epsilon (Fig. 11(b))
in terms of testing accuracy prediction. The proposed Fig. 14. Training losses vs of proposed and baseline methods for
method reported lower median values in the box plots different data distributions: non-IID based with (a) 5 IoT clients, (b) 10
IoT clients, and (c) 20 IoT clients; and IID based with (d) 5 IoT clients,
than the baseline methods for the settings mentioned above (e) 10 IoT clients and 20 IoT clients.
regarding training error prediction. It is also observed that
the proposed method reports superior performance than
the baseline method till a certain Learning Rate, as shown added during Model learning depending on this parameter.
in Fig. 11(c). Theoretically, strong privacy indicates more noise in the
r Performance Analysis with different datasets: Our pro- model, which can be ensured by applying a small value;
posed method has robustness in different datasets (MNIST, however, it costs Model accuracy. Hence, choosing an
CIFAR10, STL10, and COVID19 chest x-ray) as well. value is a trade-off between Model accuracy and privacy
We evaluate performance with recent medical datasets to preservation. Fig. 13 and Fig. 14 demonstrate the testing
justify our method’s smart healthcare system (COVID19 accuracy and training loss for different settings of the value
chest x-ray). Fig. 12 shows how different datasets are of the proposed method and baseline method. Considering
adaptable and outperform our proposed method in terms both the IID and non-IID data distribution method of
of various client numbers. the MNIST dataset, five different settings of the privacy
r Privacy Cost Analysis: The privacy cost significantly budget have been considered in this experiment 3.0, 3.5,
impacts Model construction due to the amount of noise 4.0, 4.5, and 5.0.

Authorized licensed use limited to: Chandigarh University. Downloaded on April 09,2025 at 04:16:32 UTC from IEEE Xplore. Restrictions apply.
AKTER et al.: EDGE INTELLIGENCE: FEDERATED LEARNING-BASED PRIVACY PROTECTION FRAMEWORK FOR SMART 5815

TABLE III taken into account throughout the evaluation. These constraints
OVERALL ACCURACY AND COMPUTATION TIME FOR A DIFFERENT
SETTING OF 
will be overcome in the future by developing this work.

VI. CONCLUSION
A convergent iteration-based three-fold Federated Edge Ag-
gregator architecture with differential privacy has been proposed
and developed in this paper for smart healthcare systems. It
shows edge intelligence for federated learning at the edge layer
to help a collection of health organisations protect their privacy.
After completing iterations, the edge aggregator delivers the
blended model parameters to a central aggregator. As a result, the
global aggregator cannot track crucial data about a single user.
Extensive testing has revealed that this approach provides no-
table results regarding overall learning accuracy and time com-
plexity. It also protects against global aggregator’s unauthorised
manipulation by providing extra privacy protection. The inves-
Fig. 15. Testing accuracy vs. privacy cost analysis of the proposed
tigations involve training and testing popular datasets with IID
method for different datasets (MNIST,COVID19,CIFAR10). and non-IID distributions and assessing the overall training loss
and privacy cost. A baseline technique (NbAFL method) with
no edge aggregator and a noised parameter transmitted directly
The proposed system computes the overall testing ac- to the central aggregator is used to evaluate performance at first.
curacy by a global aggregator, aggregating the Models The suggested three-fold Federated Edge Aggregator approach
from individual edge aggregator and then applying the is then tested and compared with the recent method. It is demon-
testing dataset to the final model. Table III lists the overall strated to have a much better impact on learning accuracy, con-
testing accuracy (in percentage) and computation time (in sume less computing time, have efficient data load control, and
seconds) of the proposed method under different settings provide better privacy by assessing privacy costs. Comparative
of. As can be seen, the privacy budget has less impact results show that this strategy offers a more privacy-preserving
(78% 81% accuracy) for non-IID data distribution than paradigm for the exact privacy cost as the baseline method. In
for IID data distribution (70% 80% accuracy). Although the future, A fine-grained microservices-based Federated Edge
the IID-based model takes 1.5 times more computation Aggregator will be researched to follow this study.
time than non-IID-based; however, the computation time
varies a lot for non-IID (the 1970s-2337 s) compared with
a slight variation for IID-based for different settings of. REFERENCES
r Privacy Cost Analysis for different datasets: Fig. 15 [1] L. Baghersalimi, T. Teijeiro, D. Atienza, and A. Aminifar, “Personalised
demonstrates the overall testing accuracy compared with real-time federated learning for epileptic seizure detection,” IEEE J.
Biomed. Health Informat., vol. 26, no. 2, pp. 898–909, Feb. 2021.
the value of the proposed method in different datasets [2] techjury.net, “49 stunning Internet of Things statistics [the rise of
(MNIST, COVID19 chest x-ray, CIFAR10). Five different IoT],” Accessed: Jun. 2022. [Online]. Available: https://techjury.net/blog/
settings of the privacy budget have been considered in this internet-of-things-statistics/
[3] “HIPAA Australia: The privacy act,” 1988. [Online]. Available: https://
experiment as well 3.0, 3.5, 4.0, 4.5, and 5.0. In this figure, compliancy-group.com/hipaa-australia-the-privacy-act-1988
we can see that our proposed method has met a significant [4] S. Abdullah, J. Arshad, M. M. Khan, M. Alazab, and K. Salah,
privacy paradigm. “PRISED tangle: A privacy-aware framework for smart healthcare data
sharing using IOTA tangle,” Complex Intell. Syst., pp. 1–19, 2022,
doi: 10.1007/s40747-021-00610-8.
F. Advantages and Disadvantages [5] insidebigdata.com, “Why the future of healthcare is federated AI,” Ac-
cessed: Jun. 2022. [Online]. Available: https://insidebigdata.com/2021/
Using a popular large dataset, the suggested EI-based privacy 03/16/why-the-future-of-healthcare-is-federated-ai/
preservation in FL produces promising results. Furthermore, [6] N. Bugshan, I. Khalil, N. Moustafa, and M. S. Rahman, “Privacy-
preserving microservices in industrial Internet of Things driven smart
fixed user numbers for each EA are assigned during the per- applications,” IEEE Internet Things J., early access, Jul. 21, 2021,
formance evaluation, such as 10 clients for the first EA, 5 for the doi: 10.1109/JIOT.2021.3098980.
second, and 20 for the third, demonstrating that the suggested [7] R. Wang, J. Lai, Z. Zhang, X. Li, P. Vijayakumar, and M. Karuppiah,
“Privacy-preserving federated learning for internet of medical things un-
method allows client flexibility. der edge computing,” IEEE J. Biomed. Health Informat., early access,
Adding noise to a parameter of a learned model reduces Mar. 8, 2022, doi: 10.1109/JBHI.2022.3157725.
data quality while increasing data privacy for a customer. A [8] Z. Yan, J. Wicaksana, Z. Wang, X. Yang, and K. -T. Cheng, “Variation-
aware federated learning with multi-source decentralized medical image
user’s number is not fixed in real-time. EI is only utilised for data,” IEEE J. Biomed. Health Informat., vol. 25, no. 7, pp. 2615–2628,
pre-processing in the proposed method, not for programming Jul. 2021, doi: 10.1109/JBHI.2020.3040015.
the data collecting and decision-making functions. Although [9] W. Y. B. Lim et al., “Decentralized edge intelligence: A dynamic re-
source allocation framework for hierarchical federated learning,” IEEE
it is proved that privacy assaults may be defended in an FL Trans. Parallel Distrib. Syst., vol. 33, no. 3, pp. 536–550, Mar. 2022,
framework, performance against various privacy attacks is not doi: 10.1109/TPDS.2021.3096076.

Authorized licensed use limited to: Chandigarh University. Downloaded on April 09,2025 at 04:16:32 UTC from IEEE Xplore. Restrictions apply.
5816 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 26, NO. 12, DECEMBER 2022

[10] K. Cao, Y. Cui, Z. Liu, W. Tan, and J. Weng, “Edge intelligent [28] J. Xu and F. Wang, “Federated learning for healthcare informatics,” J.
joint optimization for lifetime and latency in large-scale cyber-physical Healthcare Informat. Res., vol. 5, no. 1, 1–19, 2021.
systems,” IEEE Internet Things J., early access, Aug. 04, 2021, [29] A. Fu, X. Zhang, N. Xiong, Y. Gao, H. Wang, and J. Zhang, “VFL:
doi: 10.1109/JIOT.2021.3102421. A verifiable federated learning with privacy-preserving for Big Data in
[11] M. S. Rahman, I. Khalil, M. Atiquzzaman, and X. Yi, “Towards privacy industrial IoT,” IEEE Trans. Ind. Informat., vol. 18, no. 5, pp. 3316–3326,
preserving AI based composition framework in edge networks using May 2022, doi: 10.1109/TII.2020.3036166.
fully homomorphic encryption,” Eng. Appl. Artif. Intell., vol. 94, 2020, [30] K. Shafique, B. A. Khawaja, F. Sabir, S. Qazi, and M. Mustaqim, “In-
Art. no. 103737. ternet of Things (IoT) for next-generation smart systems: A review of
[12] W. Fang, X. Z. Wen, Y. Zheng, and M. Zhou, “A survey of Big Data security current challenges, future trends and prospects for emerging 5G-IoT
and privacy preserving,” IETE Techn. Rev., vol. 34, no. 5, pp. 544–560, scenarios,” IEEE Access, vol. 8, pp. 23022–23040, 2020, doi: 10.1109/AC-
2017. CESS.2020.2970118.
[13] Y. Xiao, Y. Li, G. Shi, and H. V. Poor, “Optimising resource- [31] M. Gheisari, Q. Pham, M. Alazab, X. Zhang, C. Fernández-Campusano,
efficiency for federated edge intelligence in IoT networks,” in and G. Srivastava, “ECA: An edge computing architecture for
Proc. Int. Conf. Wireless Commun. Signal Process., 2020, pp. 86–92, privacy-preserving in IoT-based smart city,” IEEE Access, vol. 7,
doi: 10.1109/WCSP49889.2020.9299798. pp. 155779–155786, 2019, doi: 10.1109/ACCESS.2019.2937177.
[14] M. Keshk, B. Turnbull, N. Moustafa, D. Vatsalan, and K. K. R. Choo, [32] X. Xu, C. He, Z. Xu, L. Qi, S. Wan, and M. Z. A. Bhuiyan, “Joint
“A privacy-preserving-framework-based blockchain and deep learning for optimization of offloading utility and privacy for edge computing enabled
protecting smart power networks,” IEEE Trans. Ind. Informat., vol. 16, IoT,” IEEE Internet Things J., vol. 7, no. 4, pp. 2622–2629, Apr. 2020,
no. 8, 5110–5118, Aug. 2020. doi: 10.1109/JIOT.2019.2944007.
[15] Q. Yang, Y. Liu, Y. Cheng, Y. Kang, T. Chen, and H. Yu, “Federated [33] J. Li et al., “A federated learning based privacy-preserving smart health-
learning,” Synth. Lectures Artif. Intell. Mach. Learn., vol. 13, no. 3, care system,” IEEE Trans. Ind. Informat., vol. 18, no. 3, pp. 2021–2031,
pp. 1–207, 2019. Mar. 2022, doi: 10.1109/TII.2021.3098010.
[16] M. Yamin, Y. Alsaawy, A. B. Alkhodre, A. Sen, and A. Ahmed, “An [34] Q. Wu, X. Chen, Z. Zhou, and J. Zhang, “FedHome: Cloud-edge based
innovative method for preserving privacy in Internet of Things,” Sensors, personalized federated learning for in-home health monitoring,” IEEE
vol. 19, no. 15, 2019, Art. no. 3355. Trans. Mobile Comput., vol. 21, no. 8, pp. 2818–2832, Aug. 2022,
[17] F. Sattler, S. Wiedemann, K. R. Müller, and W. Samek, “Robust and doi: 10.1109/TMC.2020.3045266.
communication-efficient federated learning from non-iid data,” IEEE [35] M. Kim, O. Günlü, and R. F. Schaefer, “Federated learning with local
Trans. Neural Netw. Learn. Syst., vol. 31, no. 9, pp. 3400–3413, differential privacy: Trade-offs between privacy, utility, and communica-
Sep. 2020. tion,” in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process., 2021,
[18] K. Wei et al., “Federated learning with differential privacy: Algorithms pp. 2650–2654, doi: 10.1109/ICASSP39728.2021.9413764.
and performance analysis,” IEEE Trans. Inf. Forensics Secur., vol. 15, [36] N. Hudson, M. J. Hossain, M. Hosseinzadeh, H. Khamfroush, M.
pp. 3454–3469, 2020, doi: 10.1109/TIFS.2020.2988575. Rahnamay-Naeini, and N. Ghani, “A framework for edge intelligent smart
[19] J. J. P. C. Rodrigues, S. Jabbar, M. Abdallah, C. Verikoukis, and M. distribution grids via federated learning,” in Proc. Int. Conf. Comput. Com-
Guizani, “Future communication trends toward Internet of Things ser- mun. Netw., 2021, pp. 1–9, doi: 10.1109/ICCCN52240.2021.9522360.
vices and applications,” IEEE Wireless Commun., vol. 26, no. 6, pp. 6–8, [37] Y. Tu, Y. Ruan, S. Wagle, C. G. Brinton, and C. Joe-Wong, “Network-
Dec. 2019, doi: 10.1109/MWC.2019.8938176. aware optimization of distributed learning for fog computing,” in Proc.
[20] X. Liu et al., “Privacy and security issues in deep learning: A sur- IEEE Conf. Comput. Commun., 2020, pp. 2509–2518, doi: 10.1109/IN-
vey,” IEEE Access, vol. 9, pp. 4566–4593, 2021, doi: 10.1109/AC- FOCOM41043.2020.9155372.
CESS.2020.3045078. [38] S. Svorobej et al., “Simulating fog and edge computing scenarios: An
[21] I. Rosenberg, A. Shabtai, Y. Elovici, and L. Rokach, “Adversarial machine overview and research challenges,” Future Internet, vol. 11, no. 3, 2019,
learning attacks and defense methods in the cyber security domain,” ACM Art. no. 55.
Comput. Surv. (CSUR), vol. 54, no. 5, pp. 1–36, 2021. [39] X. Zhang, Y. Wang, S. Lu, L. Liu, L. Xu, and W. Shi, “OpenEI: An open
[22] Y. Li, H. Li, G. Xu, T. Xiang, X. Huang, and R. Lu, “Toward secure framework for edge intelligence,” in Proc. IEEE 39th Int. Conf. Distrib.
and privacy-preserving distributed deep learning in fog-cloud computing,” Comput. Syst., 2019, pp. 1840–1851, doi: 10.1109/ICDCS.2019.00182.
IEEE Internet Things J., vol. 7, no. 12, pp. 11460–11472, Dec. 2020, [40] N. Bouacida and P. Mohapatra, “Vulnerabilities in federated learn-
doi: 10.1109/JIOT.2020.3012480. ing,” IEEE Access, vol. 9, pp. 63229–63249, 2021, doi: 10.1109/AC-
[23] Q. Wang, Y. Xiao, H. Zhu, Z. Sun, Y. Li, and X. Ge, “Towards CESS.2021.3075203.
energy-efficient federated edge intelligence for IoT networks,” in Proc. [41] H. Y. Tran and J. Hu, “Privacy-preserving Big Data analytics a compre-
IEEE 41st Int. Conf. Distrib. Comput. Syst. Workshops, 2021, pp. 55–62, hensive survey,” J. Parallel Distrib. Comput., vol. 134, pp. 207–218, 2019.
doi: 10.1109/ICDCSW53096.2021.00016. [42] S. Desai, R. Alhadad, N. Chilamkurti, and A. Mahmood, “A survey of
[24] S. Singh, R. Sulthana, T. Shewale, V. Chamola, A. Benslimane, and privacy preserving schemes in IoE enabled smart grid advanced metering
B. Sikdar, “Machine-learning-assisted security and privacy provisioning infrastructure,” Cluster Computi., vol. 22, no. 1, pp. 43–69, 2019.
for edge computing: A survey,” IEEE Internet Things J., vol. 9, no. 1, [43] M. Keshk, B. Turnbull, E. Sitnikova, D. Vatsalan, and N. Moustafa,
pp. 236–260, Jan. 2022. “Privacy-preserving schemes for safeguarding heterogeneous data sources
[25] H. K. Bharadwaj et al., “A review on the role of machine learning in cyber-physical systems,” IEEE Access, vol. 9, pp. 55077–55097, 2021,
in enabling IoT based healthcare applications,” IEEE Access, vol. 9, doi: 10.1109/ACCESS.2021.3069737.
pp. 38859–38890, 2021. [44] M. Seif, R. Tandon, and M. Li, “Wireless federated learning with lo-
[26] V. Hassija, V. Chamola, B. C. Bajpai, and S. Zeadally, “Security issues in cal differential privacy,” in Proc. IEEE Int. Symp. Inf. Theory, 2020,
implantable medical devices: Fact or fiction?,” Sustain. Cities Soc., vol. 66, pp. 2604–2609, doi: 10.1109/ISIT44484.2020.9174426.
2021, Art. no. 102552. [45] L. Huang, Y. Yin, Z. Fu, S. Zhang, H. Deng, and D. Liu, “LoAdaBoost:
[27] G. A. Kaissis, M. R. Makowski, D. Ráckert, and R. F. Braren, “Secure, Loss-based AdaBoost federated machine learning with reduced compu-
privacy-preserving and federated machine learning in medical imaging,” tational complexity on IID and non-IID intensive care data,” PLoS One,
Nature Mach. Intell., vol. 2, no. 6, pp. 305–311, Jun. 2020. vol. 15, no. 4, 2020, Art. no. e0230706.

Authorized licensed use limited to: Chandigarh University. Downloaded on April 09,2025 at 04:16:32 UTC from IEEE Xplore. Restrictions apply.