Heliyon 10 (2024) e38137

Contents lists available at ScienceDirect

Heliyon
journal homepage: www.cell.com/heliyon

Review article

Federated learning: Overview, strategies, applications, tools and

future directions
Betul Yurdem a , Murat Kuzlu b , Mehmet Kemal Gullu a , Ferhat Ozgur Catak c,∗ ,
Maliha Tabassum b
a Department of Electrical and Electronics Engineering, Izmir Bakircay University, Izmir, Turkey
b Batten College of Engineering and Technology, Old Dominion University, Norfolk, VA, USA
c Department of Electrical Engineering and Computer Science, University of Stavanger, Rogaland, Norway

A R T I C L E I N F O A B S T R A C T

Keywords: Federated learning (FL) is a distributed machine learning process, which allows multiple nodes
Data privacy to work together to train a shared model without exchanging raw data. It offers several key
Federated learning advantages, such as data privacy, security, efficiency, and scalability, by keeping data local and
Distributed machine learning
only exchanging model updates through the communication network. This review paper provides
a comprehensive overview of federated learning, including its principles, strategies, applications,
and tools along with opportunities, challenges, and future research directions. The findings of this
paper emphasize that federated learning strategies can significantly help overcome privacy and
confidentiality concerns, particularly for high-risk applications.

1. Introduction

In recent years, Artificial Intelligence (AI) has become more popular with advanced computing technologies and has been applied
in almost all sectors. For instance, in the gaming industry, when AlphaGo [1] defeated the top human Go players, it demonstrated
the huge potential of AI. More complex cutting-edge AI technology was seen in many sectors such as healthcare, finance, and many
applications. With the success of AlphaGo, a new technology called Federated Learning (FL) came into AI, providing the solution to
the problems faced earlier. Federated learning is a distributed machine learning (ML) technique that uses multiple servers to share
model updates without exchanging raw data. This helps to overcome the sensitivity to data and privacy-protected technology [2].
FL has been proposed as an alternative approach under the coordination of a central server, and a global model is trained in the form
of a federation of devices that participate [3–5]. However, in terms of data science, there are some drawbacks.
First, AI and ML models require many datasets to train them successfully. In recent years, datasets have grown larger, and models
have become more complex. Training ML models require optimization and distribution of model parameters on multiple machines
[4]. Especially with medical data, the main problem arises when a model is trained with a huge dataset. It becomes more critical
when working with medical data as the information becomes highly sensitive. There is great concern among the public, media, and
government about recent news of leaks. For example, Facebook’s recent breach of privacy caused several protests about data privacy
[6]. In response, the European Union, on 25 May 2018, enforced the General Data Protection Regulation (GDPR) [7] to strengthen

* Corresponding author.
E-mail addresses: [email protected] (B. Yurdem), [email protected] (M. Kuzlu), [email protected] (M.K. Gullu), [email protected]
(F.O. Catak), [email protected] (M. Tabassum).

https://doi.org/10.1016/j.heliyon.2024.e38137
Received 20 March 2024; Received in revised form 16 September 2024; Accepted 18 September 2024
Available online 20 September 2024
2405-8440/© 2024 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license
(http://creativecommons.org/licenses/by/4.0/).
B. Yurdem, M. Kuzlu, M.K. Gullu et al. Heliyon 10 (2024) e38137

Table 1
Total number of selected articles relevant to the topic.

Keywords Total number

“Federated learning” 3768

“Application” + “Federated learning” 3531
“Framework” + “Federated learning” 3304
“Privacy” + “Federated learning” 2956
“Strategies” + “Federated learning” 2919
“Security” + “Federated learning” 2543
“Tool” + “Federated learning” 2343
“Overview” + “Federated learning” 1918
“Aggregation” + “Federated learning” 1865
“Attack” + “Federated learning” 1832
“Heterogeneity” + “Federated learning” 1203
“Blockchain” + “Federated learning” 1254

laws to protect data security and privacy. European Parliament also regulates that users have complete power over their data [8].
People are also paying more attention to their data and where these data are used these days [9,10]. Thus, prior consent is required
to use such data.
Second, if the number of datasets is limited, there is a restriction to the development of the training model. That is, the relation
between the number of datasets and the model training is proportional. In a dataset, there is the possibility of the presence of all
kinds of scenarios. Such as datasets that help detect a chronic wound to make it possible to diagnose the disease before any severe
measures. However, obtaining this kind of data is difficult due to its sensitivity and high regulation [11].
Third, when dealing with healthcare data, a person’s data becomes highly sensitive. This type of data contains private information.
Moreover, while a large dataset is needed to train an AI model, the limited number of datasets hinders the development of ML models
[12]. Due to its sensitivity and different types of regulations, obtaining this kind of data is difficult [13].
For example, the initial version of AlphaGo relied on 160,000 sets of human chess data. This had the ability to defeat beginner-level
players. On the contrary, AlphaZero used human-generated chess data that has been shown to be capable of defeating professional
players [14]. FL-based training approaches for a neural network that includes image classification [15] or natural language process
[16] are promising along with new methods developed daily, and several new mobile applications are being developed. For example,
the prediction of Google keyboard words was improved using FL by Hard et al. [17].

1.1. Research methodology

The research process for this review paper focused on a systematic way to provide an extensive survey of the overview, strategies,
applications, tools, and future directions of Federated Learning. For this, publications and preprints consisting of review articles,
research articles, book chapters, and conference papers on academic databases consisting of Web of Science, Elsevier, IEEE Xplore,
Scopus, Springer, and Google Scholar were examined.
During the research studies, the main keyword used in searches was “federated learning”. Additionally, keywords involved in data
privacy and security, data heterogeneity, model aggregation, framework, and application topics were used in the review. Since it is a
timely and wide topic, there may be unobserved publications on this subject with the possibility that similar terms may be overlooked.
In this research, since 2016 was the first year of proposing the FL concept [4], the publications after this date were examined. Table 1
shows the number of selected articles found with the given main and additional keywords, the selection conditions of which were
previously specified.
As can be seen from the table, although it is a new field of study, there are many studies in almost every sub-branch of this
subject, and in this paper, these branches are explained. Overall, this paper is organized as follows. Section 2 provides an overview
of Federated Learning, divided into two subsections covering the characteristics and categorization of FL. Section 3 provides various
FL strategies. The security concerns associated with FL are addressed in Section 4. Section 5 presents various applications of FL.
Section 6 examines some of the FL frameworks and tools. In the last section, opportunities, challenges, and future directions in FL
are explained.

2. Overview of federated learning

Federated learning aims to build global AI models for multiple parties with diverse interests. Instead of collecting and combining
data from different locations and gathering them in a central location, FL processes data from the place where it ensures the security
of sensitive data and models.
Federated learning has undergone significant development since its introduction. Initially, researchers proposed distributed
stochastic gradient descent (SGD) as a way to train deep neural networks on multiple devices. However, this approach requires
that each device transmit its local model to a central server after every iteration, raising privacy and communication issues. In re-
sponse, Google researchers proposed the FL approach in 2016 [18], which allowed model training to occur on the devices themselves
without requiring the transmission of raw data to a central server. This approach has been extended along with various federated
learning architectures and algorithms. An extension is the use of differential privacy introduced by Google researchers [19], which

2
B. Yurdem, M. Kuzlu, M.K. Gullu et al. Heliyon 10 (2024) e38137

is a mathematical framework to quantify and guarantee privacy in the AI/ML concept. Random noise is added to the model updates
generated by participating devices in differential privacy. This approach has also been extended by using heterogeneous data [6],
where devices may have different data distributions. In addition, federated learning has been extended to other domains, such as
federated optimization [20] and federated reinforcement learning (FRL) [21]. In federated optimization, multiple entities cooperate
to optimize a shared objective function, while agents learn to act in a decentralized environment without central coordination in FRL.
These extensions will significantly contribute to the potential applications and open new research directions in federated learning.
The future of federated learning promises additional developments, including the exploration of federated learning with non-IID
data, the integration of federated learning with blockchain technology, and the development of more robust and scalable federated
learning algorithms.
In recent years, federated learning has become more mature with its characteristics and categorizations depending on how the
data is distributed, stored, and interrelated. The characteristics and categorization of FL are discussed in the following subsections.

2.1. Characteristics of federated learning

FL is a field of distributed machine learning, which puts a great deal of effort into data privacy. The latest studies also pay
attention to distributed systems that preserve privacy. It connects multiple nodes located in different locations but is connected via
a communication network. The network is under the control of the central server, and each node undertakes different parts of the
same task to complete it. The main characteristics of FL are discussed below.

• Universal cross-organization scenario: FL was originally proposed by Google as a distributed ML technology. This allows
participants to build a global model while keeping their underlying data local. The initial concept of FL is expanded to encompass
all decentralized ML techniques that preserve privacy [6].
• Massively non-identically independent distribution: In FL, data is massively widespread among a huge number of edge notes
or devices. The data available at each node might be less than the total number of nodes.
• Decentralization: It is not completely used in a technical sense. However, there is no definitive center. Each client is not
determined to be the center, but they influence the central model. Parameter servers are used as a central server that is dominant
enough to distribute the data. It works as a resource to obtain efficient collaboration [22].
• Equality of status for each node: In FL, all parties receive the same facility. This means that who has the largest number of
data has the largest mass.

2.2. Categorization of federated learning

Federated Learning can be categorized in different ways based on data distribution. In this paper, the two most popular catego-
rizations are explained, i.e., vertical federated learning and horizontal federated learning.

• Vertical Federated Learning: It is suitable in cases where data is partitioned in the verticle direction in accordance with feature
dimension. Vertical federated learning (VFL) is where data features are split among multiple parties. It is the concept that
collaborates to train a model on a dataset. Here, all parties have homogeneous data. That means that they overlap partially on
sample ID but are different in feature space.
For example, a medical organization intends to work to identify and predict diseases related to diabetes. According to research,
people with high blood pressure and obesity are likely to develop Type-2 diabetes [23]. Therefore, analyzing the weight and age
of patients can give rough dimensions along with their medical history. However, a young patient who consumes more sugar
and lacks physical activity but is not obese or has high blood pressure is more prone to suffer from diabetes. This diagnosis is not
predicted and is personalized due to lack of information. With the development of FL, some companies that work with smartphone
application datasets, such as step counters or dietary structures, can cooperate with each other. This can be done without sharing
the raw data transmissions as the figure shows. Normally, taking out similar entities that possess different characteristics to get
joint training is a path taken by scholars.
Another example is that patient data can be present in different healthcare organizations. As the identity of the patient is secret,
those two different data cannot be merged to train a model. In order to address this issue, it is necessary to train a machine
learning model through a collaborative effort that will be conducted within the appropriate premises. Unlike horizontal FL,
vertical FL poses additional challenges in terms of entity resolution [24]. Unlike the straightforward approach of aggregating all
datasets on a common server, this method is not effective in vertical FL. An example of how vertical FL works is given in Fig. 1.
It can be inferred that there is still more potential for enhancing vertical FL to be utilized in a more intricate ML approach.
• Horizontal Federated Learning: In the case of horizontal FL, similarity exists in the data features that are spread across various
nodes. Meanwhile, those data are different in sample space. The existing FL algorithms are primarily aimed at various applica-
tions of IoT devices or smart devices [2]. In these scenarios, FL is classified as horizontal-federated learning. Here, data in the
sample space may differ drastically but have similar features. Furthermore, a hierarchical heterogeneous horizontal FL frame
was introduced to meet the criteria for limited labeled entities [25]. Adapting a heterogeneous domain can solve the issue of
label shortage adopted multiple times using each participant as the target domain. In real applications, such as healthcare, data
collection is inseparable from a huge number of work. It is almost impossible for each hospital to build a data pool to share data.

3
B. Yurdem, M. Kuzlu, M.K. Gullu et al. Heliyon 10 (2024) e38137

Fig. 1. An example of vertical FL-based application [2].

Fig. 2. An example of horizontal FL-based application [2].

Therefore, in horizontal FL, multiple parties train a model with similar datasets from different sources [26]. An example of this
is illustrated in Fig. 2.

4
B. Yurdem, M. Kuzlu, M.K. Gullu et al. Heliyon 10 (2024) e38137

Table 2
Comparison of categorization of federated learning [27].

Criterions VFL HFL

Features Different Common

Samples Common Different
Training Locally on client devices
Shared with Server Local model updates
Get from server Global model updates
Data sharing with global server No
Privacy and security Data is not shared with the global server
Scalability Low High
Computing power Bounded with the device capacities of clients
Communication resources Consumes less than HFL Consumes more than VFL
Lower communication cost
Pros Huge and detailed datasets
than centralized ML
Works with a small
Cons Data heterogeneity
number of clients

In FL models, one model can have multiple characteristics discussed in the previous section. The different types vary in their needs.
The categorization is not mutually exclusive. In Table 2, the comparison of these two models is shown according to their feature and
sample diversity, training point, data sharing, privacy and security, scalability, computation features, and communication resources.

3. Federated learning strategies

In FL, multiple parties can collaborate without sharing raw data. Each party shares model updates with a central server. The data
go through aggregation by the central server and are sent to local sources. In this way, the model is trained while benefiting all parties
by providing privacy.
Distributing the data across different devices is more convenient than gathering all the data in one location or device. To achieve
this, a variety of federated learning strategies have been developed [28]. Selected ones from the commonly used ones are explained
in the next subsections.

3.1. Algorithms

3.1.1. Federated averaging

Federated Averaging, also known as FedAvg, is the most popular FL strategy. In FedAvg, a local model gets trained by each client
with their own data. Allocating parameters to a large number of clients, recycling, training, and updating them are done by the central
server while not sharing the client’s information. The whole communication is done by the central client. The model parameters across
all parties are averaged by aggregation. After that, each client receives the global model, and this process is then continued until the
model reaches convergence. This process is explained in Appendix A - Algorithm 1. This model can provide the privacy promised by
FL models. The key factor behind this is to minimize communication overhead. In the ML model, this is a promising approach. There
are some significant advantages of FedAvg over other centralized ML models:

• It provides data privacy. Data is stored in the local model while the model updates are shared. In this case, a central server is
not used to share data.
• Model updates are only shared between participants. It results in a significant amount of communication overhead, which makes
it environmentally friendly.
• It is scalable to large-scale FL scenarios with a large number of participants. The participants have the freedom to join or leave
the network at any time.

3.1.2. FedAdam
FedAdam is the reformulation of the Adam optimizer for FL. In this approach, training data is distributed among devices. The
moving average of the model update follows the momentum decay rule. This technique adaptively updates the model’s weights by
utilizing moving averages and adjusting the learning rate for each weight. The exponential moving average of the first moment of
the gradient is computed. The weights are updated with a learning rate and adapted based on the estimated variance and mean
gradient. The primary distinction between FedAdam and the standard Adam algorithm is the use of a federated average to calculate
the gradient across all devices. The model can also run well with the new data while preserving the old data.

3.1.3. FedYogi
In the case of FL, this strategy lets users train a model where the identity of the clients is not shared with the central server. This
process avoids the collection of sensitive data that can breach privacy. Mostly, this is achieved by learning a single global model. This
model is for all users, even if they are using different data distributions. For instance, users of a device prefer different suggestions
while training a model. The personalization of this heterogeneity is the motivation behind each user’s global model. In the case of
privacy, learning the full model is prohibited for privacy reasons. That is why Federated Optimization is adapted with Yogi.

5
B. Yurdem, M. Kuzlu, M.K. Gullu et al. Heliyon 10 (2024) e38137

Table 3
Comparison of FL strategies.

FL Strategy Utility

FedAvg Simple and common

FedAdam Effective optimization
FedYogi More stable learning
FedAdagrad Automatic learning rates
FedProx Better for non-IID data
Scaffold More stable convergence

3.1.4. FedAdagrad
FedAdagrad is the short form of Federated Adaptive Gradients, which is an extension of the gradient descent optimization al-
gorithm. The gradients seen for the variable, i.e., partial derivatives, are optimized and observed throughout the search. Regarding
all the model parameters, the algorithm is based on adapting the learning rate and first-order information with some properties of
second-order methods and annealing. With the growth of the dynamic rate, it is inversely proportional to the magnitude of the gra-
dient. The relationship between large gradients and smaller learning rates is proportional. The scale of gradience varies in each layer
by several orders. Moreover, the denominator of the scaling coefficient has a similar effect as annealing, which gradually decreases
the learning rate over time. The Appendix A - Algorithm 2 describes all FedAdam, FedYogi, and FedAdagrad strategies.

3.1.5. FedProx
This strategy is used to lessen statistical heterogeneity. It introduces terms that regularize each local training loss and is built on
FedAvg [29]. The local model and its deviation are also controlled by this model. Re-parameterization does some minor modifications
but it is a better version of FedAvg. It is important for both theoretical and practical fields. In terms of practical field, robust conver-
gence is demonstrated compared to FedAvg [30]. Meanwhile, in theory, in device-level systems, convergence is provided that allows
each device. Between FedProx and FedAvg, the first one is more stable and provides accurate convergence. The FedProx algorithm is
represented in Appendix A - Algorithm 3.

3.1.6. Scaffold
Stochastic Controlled Averaging for Federated Learning is also known as Scaffold. The training process is coordinated across
different nodes across the network as expected. The Scaffold algorithm is given in Appendix A - Algorithm 4. This approach reduces
the need for multiple rounds of communication and resolves the issue of heterogeneity [31]. Furthermore, a server-side learning rate
is employed to regulate variates and reduce control variates. Ultimately, it allows for collaboration between multiple devices with
different data sources.

3.2. Comparison of the FL strategies

In this section, the applications given in the literature are examined and a comparative evaluation of the FL strategies used is
accomplished.
In a study by Gao et al. [32], in addition to the FL models trained with the previously mentioned known strategies such as FedAvg,
FedProx, and Scaffold, the model trained with the FedDC strategy they proposed was also compared. Thus, the effects of these algo-
rithms for a more convergent model due to heterogeneous datasets were investigated. As a result of studies conducted with different
datasets, the FedAvg method had the most round numbers to converge to the target accuracy. Although a higher or approximate
number of rounds were obtained with FedProx, trainings were completed faster than FedAvg in some cases. With Scaffold, fewer and
faster training processes were completed compared to both methods, and the same percentage of accuracy was achieved with all of
them. The FedDC strategy, a combination of FedProx and Scaffold, showed much better results than others for both IID and non-IID
datasets. In the final comparison results, they determined that FedDC is robust and effective in full and partial client participation
[32].
Nguyen et al. conducted a study with the CIFAR-10 dataset using strategies such as FedAvg, FedProx, Scaffold, FedNova, FedAdam,
FedAdagrad, and FedYogi and their combinations [33]. In the obtained test accuracy graph, it was observed that the best results were
achieved with the ProxYogi strategy, which is the combination of FedProx and FedYogi [33]. This study was also conducted with a
non-IID dataset.
In another study, training was carried out with medical data [34]. FedAvg, FedAdam, FedYogi, FedAdagrad, FedProx, and FedAvgM
strategies were used to predict in-hospital mortality and acute kidney injury. In the results, it was given that the training with FedProx
had the lowest accuracy for the prediction of acute kidney injury. Nearly all of them obtained highly accurate training results for
predicting hospital mortality, but FedAdam had a slightly lower result. They also highlighted that the basic FL algorithms FedAvg
and FedAvgM are the best for machine learning tasks, and FedAvg performs well with the IID datasets [34].
As can be understood from the literature, there are cases where the use of each FL strategy is advantageous. An overview of these
advantages for commonly used strategies is given in Table 3.

6
B. Yurdem, M. Kuzlu, M.K. Gullu et al. Heliyon 10 (2024) e38137

4. Security concerns for federated learning

FL is a distributed machine learning approach in the AI/ML concept that involves multiple parties, which introduces various
security concerns regarding the privacy, confidentiality and integrity of the data used in the collaborative learning process. In this
section, the main security concerns associated with federated learning are discussed along with potential solutions to mitigate these
risks.

4.1. Data privacy

To mitigate the risk of data privacy, several privacy-preserving techniques have been proposed. One of these methods is differential
privacy [35], which adds random noise to the updates sent by each device to obfuscate individual contributions. For example, let 𝑋 be
the input data from a device and 𝑓 (𝑋) be the corresponding model update. Then, the noise added to 𝑓 (𝑋) is drawn from a probability
distribution 𝑁(0, 𝜎 2 ), where 𝜎 is the privacy parameter that controls the level of noise added. Mathematically, the noisy update can
be expressed as 𝑓 ′ (𝑋) = 𝑓 (𝑋) + 𝜂 , where 𝜂 ∼ 𝑁(0, 𝜎 2 ). The privacy guarantee of differential privacy can be quantified using the
concept of 𝜖 -differential privacy, which limits the amount of information that can be inferred about any individual device’s data from
the overall output of the federated learning algorithm. Specifically, a federated learning algorithm is said to be 𝜖 -differentially private
if, for any two input datasets 𝑋 and 𝑋 ′ that differ by a single record, and for any output 𝑆 of the algorithm, the following condition
holds:
𝑃 𝑟[𝑆(𝑋)]
≤ 𝑒𝜖 (1)
𝑃 𝑟[𝑆(𝑋 ′ )]
where 𝑃 𝑟[𝑆(𝑋)] and 𝑃 𝑟[𝑆(𝑋 ′ )] denote the probability of obtaining the output 𝑆 on datasets 𝑋 and 𝑋 ′ , respectively.
Another method is secure multi-party computation (SMC) [36], which enables collaborative computation on encrypted data without
revealing the raw inputs. For example, it is assumed that 𝑋1 , 𝑋2 , ..., 𝑋𝑛 present the private inputs of 𝑛 devices participating in the
federated learning process. In SMC, first, each device encrypts its input utilizing a public key, i.e., generates the encrypted data, and
then sends the data to a central server. The central server is responsible for computation on the encrypted data as well as returning
the encrypted result to the devices to decrypt the data using the private key.
The security of SMC is based on cryptographic techniques, such as homomorphic encryption (HE), which allows computations on
encrypted data. It is assumed that 𝐸𝑛𝑐(𝑋) is the encrypted version of input 𝑋 , 𝐷𝑒𝑐(𝑋) is the decryption function, and let 𝑓 be a
function operating on encrypted data. HE ensures the function can be performed on encrypted data without revealing the underlying
inputs, i.e., 𝐷𝑒𝑐(𝑓 (𝐸𝑛𝑐(𝑋1 ), 𝐸𝑛𝑐(𝑋2 ), ..., 𝐸𝑛𝑐(𝑋𝑛 ))) ≈ 𝐸𝑛𝑐(𝑓 (𝑋1 , 𝑋2 , ..., 𝑋𝑛 )).

4.2. Model poisoning attacks

FL includes model updates from multiple devices to form a global model. However, FL can be exploited by malicious participants,
i.e., injecting poisoned updates to corrupt the global model or model poisoning attacks. This can lead to incorrect predictions or
leakage of sensitive information. Several defense mechanisms have been proposed for model poisoning attacks.
One mechanism is anomaly detection methods based on statistical methods (clustering or outlier detection) or AI/ML (such as neural
networks or decision trees). The anomaly detection method is used to identify and remove malicious updates. This can be explained
mathematically as follows:

• Given 𝑈𝑡 = 𝑢1 , 𝑢2 , ..., 𝑢𝑛 : the set of updates received at time 𝑡, and 𝑓𝑡 : the global model at time 𝑡.
• Identify updates 𝑢𝑖 ∈ 𝑈𝑡 , which do not align with the current global model 𝑓𝑡 , i.e., ||𝑓𝑡 − 𝑢𝑖 || > 𝜃 , where 𝜃 is a predefined
threshold.
• Define the distance metric, i.e., ||𝑓𝑡 − 𝑢𝑖 ||, by using the Euclidean distance or the cosine similarity.
• Identify and remove malicious updates if needed.

An alternative mechanism is robust aggregation algorithms, which can be defined as an optimization problem solved using vari-
ous techniques, such as gradient descent or stochastic optimization. The robust aggregation algorithms aim to reduce the effect of
malicious updates on the global model. This can be simply expressed mathematically as follows:

• Given 𝑓𝑡 : the global model at time 𝑡, and 𝑈𝑡𝑚 = 𝑢𝑚 , 𝑢𝑚 , ..., 𝑢𝑚

1 2 𝑛𝑚 : the set of malicious device updates at time 𝑡, 𝐿(𝑓𝑡 , 𝑈𝑡 ): the loss
function of the global model in the set of updates 𝑈𝑡 , and 𝑘: a constraint that limits the maximum number of compromised
updates.
• Define the optimization problem as follows:

min 𝔼𝑈𝑡 ∖𝑈 𝑚 [𝐿(𝑓𝑡 , 𝑈𝑡 )] subject to |𝑈𝑡𝑚 | ≤ 𝑘 (2)

𝑓𝑡 𝑡

• Apply the defined optimization algorithm to the global model.

7
B. Yurdem, M. Kuzlu, M.K. Gullu et al. Heliyon 10 (2024) e38137

4.3. Model inversion and membership inference

FL can be vulnerable to model inversion and membership inference attacks. Model inversion attempts to uncover sensitive infor-
mation from the trained model, while membership inference attempts to determine if a particular data point is used in the training.
These attacks can be more critical particularly high-risk applications, such as healthcare or finance.
Fortunately, a variety of countermeasures have been developed to enhance the security of federated learning against these attacks,
which include model regularization, adversarial training, and gradient obfuscation. Each countermeasure method is briefly explained
below.

• Model regularization is an additional regularization term to the loss function during training, which penalizes the model for
learning sensitive information. This can be given mathematically as:

regularized =  + 𝜆 ⋅ R(𝑓 ) (3)

where  is the original loss function, 𝜆 is a regularization parameter, and R(𝑓 ) is a regularization term to encourage the model
to forget sensitive information.
• Adversarial training is the training process against adversarial examples generated specifically to extract sensitive information.
Mathematically, this can be given as a min-max optimization problem:

min max (𝑓 , 𝑥, 𝑥′ ) (4)

𝑓 𝑥′ ∈ ′

where 𝑥 is the original input, 𝑥′ is the adversarial example, and (𝑓 , 𝑥, 𝑥′ ) is the loss function that captures the vulnerability of
the model to model inversion attacks.
• Gradient obfuscation protects sensitive information by perturbing gradients during the training process, making it more difficult
for an adversary to perform model inversion attacks.

Additionally, a variety of countermeasures such as privacy-preserving data synthesis and differentially private aggregation can also
be utilized for membership inference attacks. Each countermeasure method is briefly explained below.

• Privacy-preserving data synthesis refers to the process of generating synthetic data that maintains the statistical characteristics
of the original data while ensuring the confidentiality of individual data points. It uses techniques such as generative adversarial
networks (GANs) or differential privacy mechanisms. Mathematically, this can be given as:

𝑋̂ = Synthesize(𝑋) (5)
where 𝑋 is the original dataset and 𝑋̂ is the synthesized dataset. The synthesized dataset 𝑋̂ can be used for training the federated
model while protecting the privacy of individual data points.
• Differentially private aggregation is to protect membership privacy by adding random noise to the aggregation process. Math-
ematically, this can be given as:
𝑛
1∑
Agg(𝑈1 , 𝑈2 , ..., 𝑈𝑛 ) = 𝑈 + noise (6)
𝑛 𝑖=1 𝑖
where 𝑈1 , 𝑈2 , ..., 𝑈𝑛 are the updates from the participating devices, and noise is a random noise term that ensures differential
privacy guarantees.

These techniques provide additional security measures to mitigate the risks associated with model inversion and membership inference
attacks in federated learning.

5. Federated learning-based applications

FL can be used in different fields, and any field that deals with a person’s private information can benefit from FL. Below are some
of the applications that can use and benefit from FL models.

5.1. Healthcare

FL shares its data without revealing the identity of the users. Each medical institute might have some data about their patient.
Still, the amount of data lacks sufficiency to train a model [37]. This makes the use of FL in this field very attractive. Some of the
applications in healthcare are discussed as follows:

• Electronic Health Records (EMR): It has a lot of clinical components, the authors [38] suggested an attempt to use tensor
factorization models. It would do phenotype analysis and obtain concealed medical data of patients. It is the first attempt at FL
learning. Their proposed graph-based attention model was used in disease diagnosis prediction with the records of adult patients

8
B. Yurdem, M. Kuzlu, M.K. Gullu et al. Heliyon 10 (2024) e38137

Fig. 3. Applications of federated learning with different IoT devices.

and intensive care unit patients, and heart failure prediction studies. The model they proposed as a result of studies conducted
with the data of patients between the ages of 20-40 in intensive care disease prediction ensured them to achieve 10% higher
accuracy than training with RNN [38]. The authors [39] explored EMR differently while using Federated Settings. It was used
to predict the mortality rate of cardiac patients by using EMRs from different hospitals.
• Biomedical Image Analysis: FL has been used in this sector. For example, Silva et al. [40] put forward FPCA (Federated principal
components analysis). It extracted features from MRI that come from different medical centers.
• EEG Classification: A Hierarchical Heterogeneous Horizontal FL (HHHFL) was proposed by Gao et al. [25]. They overcame the
challenge of limited labeled samples along with the provision of privacy [2].
• Clinical Prediction: To predict disease, beforehand FL can be used. As described by the authors Pfohl et al. [39], FL establishes
efficacy over centralized and local data. It also performs FL differently.

Overall, FL can be used to provide effective ways in healthcare that will help to make life easier.

5.2. Internet of Things (IoT)

Federated learning has several potential applications in the IoT industry. Fig. 3 shows various applications of FL and how different
IoT devices take part in training an FL model. Here are some of the potential applications of federated learning in IoT:

• Smart Transportation: Intelligent Transportation System (ITS) has been implemented by using AI/ML for the past few years. It is
done using centralized vehicular data learned from the data center [41]. This requires sharing data in an untrusted environment,
which is a safety issue for FL and is used to provide privacy and safety [42].
• Smart City: The use of AI has been increasing over the years. The creation of a smart city that can enhance the quality of life
for citizens living in urban areas is needed. By supplying seamless delivery of food, water, and other necessary items to their
users [43]. To create a smart city, many sensors are needed. Also, AI and ML have the ability to take care of huge numbers of

9
B. Yurdem, M. Kuzlu, M.K. Gullu et al. Heliyon 10 (2024) e38137

big data that are usually generated from sensors [44]. FL offers a more attractive and effective way by enabling decentralized
methods. For example, an environmental monitoring framework based on fog computing was proposed in study [45]. The studied
multi-source heterogeneous dataset was collected from the IoT sensors in Beijing.
• Smart Grid: Smart grid is an important factor in terms of increasing industrial systems as it provides the necessary power to
the energy sources of the household and industry [46]. Smart grid operations can be helped by FL with intelligent solutions. For
example, FL can be used to create a federated predictive power scheme that can run a recent neural network locally. That can
estimate the amount of power demands by inspecting the history of a customer’s usage. This can help construct a global prediction
model [47]. In the research [47], a power consumption prediction model for GPU-based edge data centers was constructed.
Their model is based on ANN and achieves a normalized root mean square deviation error of less than 7.4% compared to actual
measurements.

The FL model can help in the IoT sector in different ways. The use of different technological products is increasing day by day, and
this requires privacy, too. FL can provide this by leveraging IoT products.

5.3. Internet of Underwater Things (IoUT)

IoUT stands for Internet of Underwater Things and has garnered rapid momentum recently [48]. The use of FL in IoT networks
has been expanding rapidly in recent years. It spans applications on defense, environmental monitoring, exploration of the sea, etc.
Like any other FL system, IoUT also uses a traditional ML model. Below, the various applications of IoUT are discussed.

• Environmental Monitoring: The environment consists of different elements like water, air, soil, etc. In terms of them, water is
an essential one, and it is used in everyday life. The quality of the water always needs to be monitored because contaminated
water can cause some serious problems. FL can be used to monitor the quality of water at all times. So that, no man-made
pollution happens, like contaminating the water due to oil spills.
• Underwater Exploration: This helps to discover lost treasures such as underwater ruins of civilization and other natural re-
sources. The underwater environment is used to experiment with some new technology. How these technologies are affected is
also investigated using several sensors. H. Zhao et al. [49] proposed to use a “federated meta-learning enhanced acoustic radio
cooperative framework”, called as ARC/FML, for data collecting from distributed sources. This stated technique helps in sharing
sensitive data regarding underwater exploration. DeepSink and RF Channel were used to conduct the experiment. The proposed
model provided an accuracy of 97%.
• Disaster Prevention and Mitigation: The underwater is always susceptible to various disasters. They can be both natural and
man-made. Disasters that are man-made are mostly oil spills, and poisonous chemical spills which mostly occur due to underwater
experiments. On the other hand, natural disasters are tsunamis, earthquakes, and underwater volcano eruptions. The Deepwater
Horizon incident in 2010 [50] is one of the worst man-made disasters to be recorded. It occurred due to the leakage of a significant
amount of natural gas, which resulted in the gas rising and catching fire [51]. On the other hand, the most horrific natural disaster
occurred in 2004, which was the earthquake and tsunami in the Indian Ocean [51]. FL-enabled IoUT solutions can significantly
provide help in preventing these disasters by sending the data in real-time.
• Defenced Military: Defensive naval operations traditionally involve human-operated underwater vessels to carry out tasks such
as detecting submarines, mine warfare, recovery operations, and surveillance. However, advancements in technology have led
to the development of Underwater Wireless Sensors, or UWSN, which are a type of wireless sensor network that allows for
underwater activities to be carried out without human intervention. UWSN are capable of detecting and classifying objects in
the underwater environment. The US Navy places great emphasis on real-time data sharing between ships, submarines, drones,
and intelligence analysts onshore [52]. FL is used for this purpose, too.

These are some of the applications of FL in terms of IoUT. There are many more ways that FL can be used.

5.4. Industrial engineering

FL has several application sectors in different industries, and some of them are mentioned below.

• 6G Industry: In the recent growth in data traffic, ML has garnered a huge amount of attention. In the development of the sixth
generation (6G), it has become vital [53].
• Image Detection: FL can also be applied to perform visual inspection task [54]. It can be used to help detect defective products
in the manufacturing industry.
• Image Representation: In image field, Liu et al. [55] describes about how vision-and-language as a flashpoint. They also discuss
how FL can diversify these tasks for better grounding.
• Unmanned Aerial Vehicles: The work of unmanned aerial vehicles by using FL is given by Mowla et al. [56]. It was discussed
how these vehicles can detect malicious attacks by using FL.
• Electrical Vehicles: Electric vehicles are getting popularized now. An FL is designed by Saputra et al. [57] that can predict the
vehicles’ energy demand and methods of efficient charging stations. This can prevent energy congestion in transmissions.

10
B. Yurdem, M. Kuzlu, M.K. Gullu et al. Heliyon 10 (2024) e38137

• Financial Field: The use of credit cards and transactions in various fields is increasing day by day. Yang et al. [58] leveraged
FL to detect credit card fraud efficiently. This is a significant contribution to the field of finance.
• Text Mining: An industrial grade federated framework was used by Wang et al. [59] on Latent Dirichlet Allocation. It is used to
assess real data for spam filtering.

5.5. Integrating FL with technologies

FL is used in studies in different fields and with different data, as well as in integration with different technologies.

• Blockchain: Blockchain-supported FL applications are being implemented to solve problems, such as centralized processing, data
manipulation, and lack of motivation [60]. With blockchain technology, improvement of the security and scalability of FL is also
being realized [42]. Singh et al. [61] proposed a framework to use in the IoT healthcare field, taking advantage of the unanimity
efficiency improvement of FL and the stability, originality, transparency, distributed, and decentralized features of blockchain.
It has been stated that although the study was developed only as a theoretical model, it is planned to focus on delay and storage
needs in the future [61]. BLADE-FL, a blockchain-based decentralized FL architecture, was proposed in [62]. The standard FL
frameworks work with a single central server. Hence, a malicious client may cause the training to fail. In BLADE-FL, FL and
blockchain are integrated into each client for training and mining tasks, respectively. Studies based on blockchain-based FL were
overviewed in [63]. It was concluded from the literature that integrating blockchain with FL increases security in training, but
is not sufficient to protect privacy.
• Edge Computing: Since FL is a collaborative learning and model optimization method, it can be an encouraging factor for edge
computing technology that extends computing services closer to the dataset [64]. Poisoning attacks, an attack model using edge
computing in FL, were designed by Zhang et al. [65]. Updating the global model parameters by repeating the samples of the
clients, increasing the applicability of the attacks by reducing the attack assumptions, and poisoning attack strategies such as
label flipping and backdoor were implemented. As a result, it was observed that these attacks effectively risked the global model
[65]. Ye et al. [66] proposed the EdgeFed algorithm to optimize the high computational cost encountered on mobile devices
during the implementation of FL based on edge computing with the FedAvg strategy. In this algorithm, the process of updating
the local model is separated, and the outputs of mobile devices are collected at the edge server to reduce the frequency of global
communication [66]. Another study for FL model aggregation of image classification was conducted in the field of vehicular
edge computing [67]. The model was selected using the two-dimension contract theory as a distributed framework. As a result
of studies involving the MNIST dataset, model aggregation showed better results in accuracy and efficiency than the original
FedAvg strategy [67].

6. FL-based frameworks and tools

In FL research, we find numerous libraries, many of which are open-source, that act as both frameworks and tools. They assist
in training models, ensuring security, facilitating communication, and quickly compiling results. Using these libraries helps reduce
repetitive tasks in the code, leading to more streamlined models. With the growing variety of these tools, their capabilities also
expand. Therefore, choosing the right tool or framework becomes essential for better model outcomes.

6.1. Existing frameworks and tools

Various software tools tailored for Federated Learning (FL) are emerging, each with distinct methods and capabilities. The liter-
ature showcases several notable frameworks and tools, including but not limited to: TensorFlow Federated (TFF) [68], PySyft [69],
NVIDIA Federated Learning Application Runtime Environment (NVFlare) [70], Federated AI Technology Enabler (FATE) [71], Flower
[72], IBM Federated Learning (IBM FL) [73], FedLab [74], FedML [75], Federated Learning Utilities and Tools for Experimentation
(FLUTE) [76], Open Federated Learning (OpenFL) [77], SecureBoost [78], Interplanetary File System (IPFS) [79], Fed-BioMed [80],
and FederatedScope [81]. Subsequent subsections focus on the unique features of these frequently mentioned frameworks.

• TFF: TensorFlow Federated (TFF) is an open-source framework introduced by Google, supplying decentralized data to ML [68].
It promotes global model training, leveraging local data without requiring server uploads. The results from individual devices
are then centralized for consolidation. Being rooted in TensorFlow, TFF is able for mathematical tasks as well as learning [82].
It offers the FL API and Federated Core (FC) APIs, enhancing the flexibility in developing and applying FL algorithms. While it
allows the use of Regression and Neural Network (NN) models in horizontal FL, vertical FL is not supported [83]. Unlike many,
TFF is designed for single-host deployments, emphasizing its utility where data privacy and security are paramount. The TFF
architecture in Fig. 4 includes a comprehensive framework designed to support federative learning scenarios. This architecture
includes optimization strategies such as FedAvg and FedSGD, differential privacy with security features, various model types,
and several protocols between an executor and client that enable communication at run-time.
• PySyft: Developed by the OpenMined community, PySyft is an open-source framework that prioritizes data privacy and security.
It employs methods, such as anonymization, encryption, and differential privacy to partition data and models [69]. Since it per-
forms data anonymization, this framework is used in encrypted privacy-preserving deep learning studies. The differential privacy

11
B. Yurdem, M. Kuzlu, M.K. Gullu et al. Heliyon 10 (2024) e38137

Fig. 4. Architecture of TFF [82].

Fig. 5. Architecture of PySyft [82].

techniques included in PySyft help prevent attacks against FL models that may be endangered despite being encrypted, by confus-
ing steps that increase the complexity of the training process [84]. During remote data operations, it facilitates learning without
direct data copying, maintaining privacy boundaries. Moreover, PySyft is compatible with established ML frameworks like Ten-
sorFlow and PyTorch [69]. The image describing the PySyft architecture (Fig. 5) contains the main structure of this framework. In
this structure, there are special modules to process horizontally and vertically split data and different techniques used to ensure
computational, security and confidentiality. PySyft offers a powerful framework, supporting a wide range of features includ-
ing secure multi-stakeholder computations, different data partitioning strategies, encryption techniques, and privacy-preserving
federative learning.
• NVFlare: Designed primarily for collaborative computing, this framework ensures a secure and confidential environment for
researchers [85] [70]. It operates using Python. During model training, clients independently train using their datasets and
subsequently transmit results to a central server. The server model is updated with this information, and this iterative process
persists until the main model convergence [70]. Fig. 6, which shows the NVFlare architecture in detail, includes a wide range
of features under the headings of learning algorithms, Federation workflows, FLARE programming API, privacy protection with
security management, tools, and FL simulator.
• FATE: FATE is an algorithm-centric framework focusing on machine privacy protection, with capabilities in both simulation and
federated modes [82]. It operates under the assumption of a semi-trusted server, prioritizing combined parameters over private
data [71]. Designed for AI and big data industrial applications, FATE finds usage in sectors like finance and healthcare [87] [88].
The framework is versatile, supporting NN, decision trees, regression, and transfer learning [71]. Fig. 7, which shows the FATE
framework, reveals in detail a broad structure covering major components such as cloud-based services, workflow management,
model distribution, federated ML components, security protocols, computational resources, and storage.
• Flower: The Flower FL framework is friendly, easy to use, adaptable to FL models, can be used in similar studies, and can be used
to run these studies on lots of heterogeneous devices [72]. Flower has built-in algorithms such as FedAvg, FedProx, QFedAvg,
and FedOptim, addressing challenges such as client connectivity issues and adapting to varied network conditions. These core
strategies also serve as foundations for deriving algorithms for additional functionalities [72]. The flow in Fig. 8 demonstrates in

12
B. Yurdem, M. Kuzlu, M.K. Gullu et al. Heliyon 10 (2024) e38137

Fig. 6. Architecture of NVFlare [86].

Fig. 7. Architecture of FATE [89].

detail the interaction between the core components of the Flower framework: Strategy, Client Manager and Remote Procedure
Call (RPC) Server and RPC Clients, revealing a comprehensive design that shows the workings of Flower’s internal structure.
• IBM FL: IBM FL, an independent ML library, supports Python and is adaptable for use with decision trees, support vector machines
(SVMs), NNs, and RL [73] [82]. Since this FL was developed by IBM company, it is a proprietary framework, distinct from open-
source options [82]. FL algorithms are executed by easily configuring all the basic substructures in their infrastructure, and
implementation of grouping and aggregation of these algorithms are provided. Its capacity to collate data from various sources
enhances its viability for use in institutional areas [73]. Notably, it incorporates FL strategies such as FedAvg and iterative
average, predominantly for NN and RL applications [82]. Fig. 9 demonstrates the IBM FL architecture in detail. Data sources of
different countries are connected to each country’s own Remote Training System (RTS), and it is shown how the model parameters
obtained from here are integrated and used within this framework. This flow clearly describes how data sources are integrated
and how model updates are coordinated by ensuring secure cross-country collaboration.
• FedLab: It is an open-source framework that allows model optimization, data partition, and communication between algorithms
and their efficiency can be examined with FL simulations [74]. Because of the flexible structure of the FedLab framework, users

13
B. Yurdem, M. Kuzlu, M.K. Gullu et al. Heliyon 10 (2024) e38137

Fig. 8. Architecture of Flower [90].

Fig. 9. Architecture of IBM FL [91].

can construct their own FL simulation algorithms by modifying the substructures. Three different simulation plans have been
added, namely Standalone, Cross-process, and Hierarchical [74]. These modes are used to simulate a single operation by multiple
clients with limited possibilities, to process multiple large FL models in the same environment, and to provide communication
between local and global servers, respectively [74]. The communication contract definition and FL optimization stages are
performed during the model development process. Fig. 10 shows the FedLab architecture with the interaction between the server

14
B. Yurdem, M. Kuzlu, M.K. Gullu et al. Heliyon 10 (2024) e38137

Fig. 10. Architecture of FedLab [74].

Fig. 11. Architecture of FedML [75].

and the client via the FedLab Communication API. Focusing on FedLab’s communication agreement, optimization strategies, and
computation steps, the flowchart of the framework is expressed.
• FedML: FedML is an open-source framework that offers flexibility in training. It supports different computing methods, in-
cluding distributed computing, on-device training for edge devices, and single-machine simulation [75]. It provides FL training
by supporting various algorithmic operations. It consists of two components, FedML-API, which represents the high-level API
and enables the implementation of new algorithms in accordance with the client-oriented programming interface, and FedML-
core, which represents the low-level API and divides communication and training modules [75]. The design objective behind
the low-level APIs is to amplify security, privacy, and resilience. The architecture of the FedML framework consists of FedML
Server, FedML-API, and FedML-core structures (as shown in Fig. 11). FedML Server coordinates FL processes, FedML-API per-
forms optimization operations, and FedML-core manages the framework’s communication and training operations. Integrated

15
B. Yurdem, M. Kuzlu, M.K. Gullu et al. Heliyon 10 (2024) e38137

Fig. 12. Architecture of FLUTE [92].

Fig. 13. Architecture of OpenFL [77].

work between and within these structures ensures that the framework provides training in a powerful and efficient manner.
These dynamic structures in the architecture of the FedML framework are designed to optimize the learning process and ensure
effective communication between different components.
• FLUTE: This framework was developed by the Microsoft Research team to ensure enhanced performance in FL research [76].
Tailored for FL simulations, it aims to design a framework that works better in security, optimization, and communication scales
[76]. Training procedures can be changed according to the customer’s request, and they can accommodate complex algorithms.
During the training phase, only gradients are sent to the centralized server, and the client or server can change the gradient values
for privacy [76]. FedAvg strategy is used as the basis for federated optimization. Furthermore, it seamlessly integrates with the
NVIDIA Collective Communications Library, presenting potential advantages in temporal efficiency and memory consumption
[76]. Fig. 12 is a diagram describing the architecture of the FLUTE framework. Here, it is shown how the model on the server is
updated, how the data from the clients are integrated, and how the federative approach is applied in the model training process.

• OpenFL: OpenFL framework was initially developed for applications in healthcare but has subsequently been updated for broader
industrial usage, primarily due to its robust security features [77]. Throughout the model training phase, the model owner can
only access model weight updates and measurements. The data used here is retained at the collaborating center [77]. The FL
plan and model code can be accessed prior to training with an OpenFL command [77]. It also offers developers the ability to
easily work with remote data. However, it doesn’t provide differential privacy. Fig. 13 gives the architecture of the OpenFL
framework. Here, the flow of model parameters between the collaborator and the aggregator ensures the training process. In
addition, the local file system, FL backend, and FL plan executor, which are included in the internal structure of both collaborator

16
B. Yurdem, M. Kuzlu, M.K. Gullu et al. Heliyon 10 (2024) e38137

Table 4
Comparison of the properties of the frameworks [93].

Properties
Framework Developer Open Data Usability2 Integration3 Industry Safety-focused Scalability
source sharing1 applications design

TFF Google Yes H 3 T Yes Yes Yes

PySyft OpenMined Yes H&V 3 K, P, T Yes Yes Yes
NVFlare NVIDIA Yes H&V 2 P, T Yes Yes Yes
FATE WeBank Yes H&V 3 K, P, T Yes Yes Yes
Flower Adap gmbh Yes H 1 K, P, T Yes Yes Yes
IBM FL IBM No H 4 K, P, T Yes Yes Yes
FedLab FedLab Yes H 3 K, P, T Yes Yes Yes
FedML FedML Inc. Yes H 5 J, M, P, T Yes Yes Yes
FLUTE Microsoft Yes H 2 P, T Yes Yes Yes
OpenFL Intel Yes H 5 P, T Yes Yes Yes
1
H = Horizontal Federated Learning, V = Vertical Federated Learning.
2
The ease of usage rating is high if the framework has paper, GitHub, website, and video tutorials for its use.
3
J = Jax, K = Keras, M = MXNet, P = PyTorch, T = TensorFlow.

and aggregator, constitute the basic infrastructure components of this framework. The workflows within these infrastructures
are critical elements that support the general training mechanism of OpenFL.

6.2. Criteria to choose framework

As can be seen from the frames given in the previous section, each frame has its own pros and cons. These features allow users
to make different choices according to the FL model they will train. In Table 4, comparisons of these frameworks in terms of various
features are given. For example, one of the most essential features that users can benefit from is whether the framework is open
source or not. When this matter is examined, it is seen that the sample frameworks given here, except the IBM framework, are open
source. As the next feature, when the data sharing permission is investigated, it is stated that many frameworks do not allow vertical
FL applications. The ease of use of these frameworks is rated according to the level of availability of manuals such as paper, GitHub,
website, and video tutorials. FedML and OpenFL are highly rated for having a large number of manuals. In terms of integration, it has
been observed that while TFF can be integrated into the least environment, FedML can be integrated into the most. It is understood
from the table that the last three features are taken into account in all frameworks. These can be interpreted as follows. Applications
can be implemented with these frameworks in the industrial field. The designs of the models are formed to secure privacy. All of
them have a scalable structure.

7. Opportunities, challenges, and future directions

This section provides an overview of the opportunities, challenges, and future directions of federated learning. Federated learning
is a decentralized learning model approach offering improved privacy protection while allowing the collective intelligence of various
devices or data resources. However, there are still issues to be addressed, such as communication, system and data heterogeneity,
cyber-attacks, etc. Future research and direction in Federated Learning will primarily focus on these challenges.

7.1. Opportunities

Federated learning is one of the promising fields in the AI/ML concept. It also offers many opportunities. The primary ones can
be summarized below.

• Preserving privacy: Training provided with federated learning is carried out on the local devices of the clients. Only the updated
parameters of the model are shared with the server, and the data used in training does not need to be shared [94]. With this
method, the confidentiality of the data can be ensured. To provide privacy secure aggregation protocols can be introduced that
devise a training framework that can protect the local model updates [95] [96]. These protocols can enable the server to create
a combination between the global model’s update and each individual local model without learning about any information. This
prevents the local model updates from staying concealed from the server which prevents the server from exploiting the updates
of any user to infer their private training data [97].
• Improved data security: The data does not leave the device and is processed locally. The information shared with the server is
limited to only updated model parameters. Therefore, the server does not have access to any client’s own training data.
• Scalability: During the FL process, model training can be carried out in parallel on different devices of the clients. Here, the
processed dataset can be divided into users, and its size can be reduced. Additionally, the location of the devices does not matter
as long as they have a connection to the server. As a result of the convenience brought by such examples, it can be said that
model training with FL is scalable.

17
B. Yurdem, M. Kuzlu, M.K. Gullu et al. Heliyon 10 (2024) e38137

• Reduced communication costs: Especially for AI training, the processed datasets must be quite large. Hence, the accuracy of the
training increases. In addition, in studies developed collaboratively, a large dataset can be obtained by combining the datasets of
different clients. In studies conducted with FL, working with a large amount of data requires less communication cost. Because
each client processes its own dataset and shares only the training parameters with the server without sharing the dataset. For
this, low bandwidth is sufficient [64].
• Edge computing integration: With Edge Computing (EC), IoT data obtained from devices such as sensors is used in studies
[64]. However, this data is directly shared with third parties. Since the data is shared, communication costs are high. Data
confidentiality also cannot be ensured. However, it is possible to integrate this architecture with FL. Thus, EC studies can be
performed using the opportunities provided by FL.
• Real-time adaptability: As aforementioned, model training is done on local devices. The model parameters obtained as a result
of the training are instantly shared with the server. Therefore, clients can directly access the current model. In this way, efficient
and fast training processes are carried out in real-time.
• Decreased data bias: In the training performed by different clients on their own devices, data that do not have the same content
and type are used. These data can be IID (independent and identically distributed) or non-IID. Thus, data diversity increases.
This increases the generalization performance of the updated model and reduces data bias.

7.2. Challenges

Research on FL is still in its primary stages. Federated Learning has several challenges. Some of them are discussed below.

• Security and Privacy Challenges: The fundamental premise of Federated Learning is to provide privacy to the local datasets. A
secure aggregation algorithm is proposed that can aggregate encrypted local models without decrypting the data in the aggregator
[98]. However, a specific local learner’s identity can be disclosed. It can be done by analyzing the global model [99]. To prevent
this, differentially private federated learning has been proposed [100]. Necessary privacy is provided at the local learner level
instead of providing protection to a single data sample.
• Wireless Communication and Settings: The other challenge is wireless communication (widely used in real-world applications)
due to limited channel capacity, noise, and interference. The information is quantized before being sent over to the channels.
This is done by exchanging model parameters between the local learners and the aggregator. The federated learning paradigm
takes off with parameter quantization [101].
• Communication Overloads: In federated learning, it is one of the major challenges. Existing studies tried to solve this issue by
applying data compression [4] or by allowing clients to add only relevant output [102,103].
• System and Data Heterogeneity: In the FL, the system and data heterogeneity in the network along with the non-identically
distributed data from the nodes significantly affect the overall model and system performance. An FL model trained on hetero-
geneous, i.e., non-IID datasets may bias training in the direction that causes heterogeneity of datasets stored locally by clients.
As observed in the literature, the problem of data heterogeneity is tried to be overcome by developed strategies used in FL. For
example, while FedAvg is more suitable for training with IID datasets, the developed FedProx, FedYogi, Scaffold, etc. strategies
were improved to adapt the strategies to non-IID datasets [32–34].
• Membership Inference Attacks: In this case, the raw data stays in the local device. Even with this step, there are several ways,
which can infer the training data used in FL. It is possible to extract the information regarding the training data [104]. It looks
for mechanisms that offer a differential privacy guarantee. Some of the defense mechanisms are discussed below.
– Secure Computation: There are two main techniques that fall under this category, i.e., Secure Multiparty Computation (SMC)
and Homomorphic Encryption (HE). In SMC, with the inputs provided by the participants only two or more parties comes
to an agreement to perform. Also, the output is only revealed exclusively to a subset of participants. In HE, computation is
performed on encrypted inputs without decrypting it first [104].
– Differential Privacy: In this scheme, before model aggregation, noise is added to mask a user’s contribution. [100].
– Trusted Execution Environment (TEE): It is a secure platform, which runs a process when provided with low computational
overhead. This is executed when it is compared with a secure computation technique.
• Data Poisoning Attacks: It is widely recognized as the most prevalent form of attack against ML models. In the context of the
FL model, this attack is carried out during the training process by introducing malicious behavior to a subset of participating
devices. This results in the compromisation of model accuracy. The adversary has the ability to directly inject poisoned data into
the devices or inject poisoned data through other devices [105]. Identifying malicious participants is the defense mechanism
against this kind of attack. It is executed in each round of learning before averaging.
• Model Poisoning Attacks: These attacks are similar to data poisoning attacks. In this case, the main objective is to corrupt the
local models rather than manipulating the local data. The global model is affected by introducing errors. The attacker carries
out a model poisoning attack by compromising certain devices and altering the parameters of the local model. The defense
mechanism is almost similar to data poisoning attacks. The most common defense mechanisms are based on rejections. It is
mostly based on two factors one is error rate and another one is loss function [106]. To overcome this challenge, Bagdasaryan et
al. have developed a model-replacement technique that is accomplished by injecting backdoors [107]. However, the persistence
and stealth of this methodology are not good enough [108]. Thus, stealth was tried to be ensured with another methodology
developed in [109]. The method was improved with the use of an alternating minimization strategy. Additionally, they noted
that their methodology is very vulnerable to Byzantine-resilient aggregation strategies [109]. In the study [108], an attack based

18
B. Yurdem, M. Kuzlu, M.K. Gullu et al. Heliyon 10 (2024) e38137

on optimizing stealthy and persistent was proposed. This was developed by bypassing defense methods and avoiding forgetting,
respectively.
• Backdoor Attacks: Secured Federated Learning provides the devices anonymity during the process of the model update. Backdoor
Attacks use the same functionality, a device or a group of devices introduces an adversary in the global model [107]. An attacker
can manipulate certain tasks without impacting the overall accuracy of the global model. This can be done by assigning a
particular label to a data instance with specific characteristics, which is commonly referred to as a targeted attack.

7.3. Future directions

In federated learning, there are many open research directions required to explore, which can address the challenges discussed
earlier. They can be briefly explained in the following.

• Wireless Settings: One of the important considerations in paradigm quantization is the robustness of the models present in the
quantization error. It also includes communication bandwidth, noise, and interference. Robustness to these channel effects might
be a consideration [101].
• System and Data Heterogeneity: FedAvg was introduced as a method to tackle this challenge. However, due to the large differences
in the structure of the datasets used for different applications, this method is not efficient in overcoming this. Modifications in
the model aggregation methods are addressed to solve this issue [110,111].
• Incentive Mechanism: FL methods operate under the assumption that devices will collaborate during the learning process as
needed, without taking rewards into account. On the contrary, nodes need to be economically compensated for their participation.
This reputation-based incentive method, where devices get rewarded according to their model accuracy, data reliability, and
contribution, might help get better models [112,113].
• Federated Learning as a Service: Federated Learning as a cloud service is recommended for collaborating among third parties.
This framework allows the third party applications to contribute and collaborate on an ML model. This was developed in a recent
work [114]. For any operating environment, the framework claims to be suitable.
• Asynchronous Federated Learning: Current FL aggregation techniques are for devices that has the ability to work in a synchronized
manner. However, due to factors such as systems and data heterogeneity along with training and model transfer a synchronized
manner is followed. In this manner, the feasibility to scale federated optimization has a chance of decreasing. This can happen
in a synchronized manner [115]. Asynchronous federated averaging techniques have the capability to accommodate a larger
number of devices, enabling updates to be received at different times in comparison to FedAvg.
• Blockchain in FL: In order to enhance the global model’s capability to handle the asynchronous arrival of device parameters, the
inclusion of an aggregator is essential. The presence of this component serves as a requirement for the widespread adoption of
FL models. Blockchain is an example of a decentralized network. This means devices have the ability to learn collaboratively
without using the central aggregator. Federated Learning was proposed by some works in blockchain framework [116].
• Backdoor Attacks: The model remains vulnerable to backdoor attacks by attackers placing backdoor triggers on local models
during the training stage. Then, at the prediction stage, attackers triggered by crafted inputs cause misclassification [117]. To
overcome this challenge, various backdoor defense methods are being developed [107].
• Poisoning Attacks: It has been determined that FL frameworks have vulnerabilities in active attacks. One of these, the poisoning
attack, occurs when attackers damage the global model with local updates prepared by attackers [118]. In addition, data poi-
soning attacks are sometimes performed by injecting malicious data into the training dataset before the learning stage begins
[119].
• Adversarial Example Attacks: In this type of attack, malicious output is produced by adding little disturbances to the input
samples [120].

8. Conclusion

This survey paper provides a comprehensive overview of Federated Learning (FL), i.e., a distributed machine learning approach,
which enables collaborative training of a shared model without sharing raw data. Unlike traditional approaches, FL facilitates collab-
orative model training without the insecure exchange of raw data, thereby safeguarding critical data privacy and security. FL offers a
solution by allowing data to remain decentralized while facilitating model training through collaborative efforts and various strate-
gies of FL. Security concerns related to FL are addressed in the survey, recognizing the importance of safeguarding data and models
in distributed learning environments. Furthermore, the survey explores the FL-based applications, revealing their transformative po-
tential across diverse domain(s). Additionally, the survey recognizes both the opportunities and challenges associated with working
with FL, underscoring the potential for enhanced machine-learning applications while acknowledging the complexities involved. It is
expected that the survey paper will contribute to the advancement and adoption of this innovative approach in the field of federated
learning in terms of data privacy, security, and scalability.

CRediT authorship contribution statement

Betul Yurdem: Writing – review & editing, Writing – original draft, Visualization, Validation, Methodology, Investigation, For-
mal analysis, Conceptualization. Murat Kuzlu: Writing – review & editing, Writing – original draft, Supervision, Data curation,

19
B. Yurdem, M. Kuzlu, M.K. Gullu et al. Heliyon 10 (2024) e38137

Conceptualization. Mehmet Kemal Gullu: Writing – review & editing, Writing – original draft, Supervision, Project administration,
Formal analysis, Conceptualization. Ferhat Ozgur Catak: Writing – review & editing, Writing – original draft, Supervision, Investi-
gation, Formal analysis, Conceptualization. Maliha Tabassum: Writing – review & editing, Writing – original draft, Formal analysis,
Conceptualization.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to
influence the work reported in this paper.

Data availability statement

Not applicable.

Appendix A. Federated learning strategies - algorithms

Algorithm 1 FedAvg [18].

1: Server performs:
2: k is indexed as the K clients,
3: the minibatch size is denoted by B,
4: the number of local epochs is E,
5: the learning rate is denoted as 𝜂
6: 𝜔0 is initialized by the server
7: for 𝑡 = 1, 2, … , 𝑇 do
8: 𝑚 ← max(𝐶.𝐾, 1)
9: 𝑆𝑡 ← 𝑚 clients (random set)
10: for 𝑘 ∈ 𝑆𝑡 do
11: 𝜔𝑘𝑡+1 ← ClientUpdate(𝑘, 𝜔𝑡 )
∑𝐾 𝑛
12: 𝜔𝑡+1 ← 𝑘=1 𝑛𝑘 𝜔𝑘𝑡+1
13: ClientUpdate(𝑘, 𝜔):
14: 𝛽 ← split 𝑃𝑘 into batches of size 𝐵
15: for each local epoch 𝑖 from 1 to 𝐸 do
16: for batch 𝑏 ∈ 𝛽 do
17: 𝜔 ← 𝜔 − 𝜂∇𝑙(𝜔; 𝑏) ⊳ 𝑙 represents the loss function
18: end for
19: end for
20: end for
21: end for
22: return 𝜔 to server ⊳ The updated model 𝜔 is returned to the server for the next iteration

Algorithm 2 FedYogi, FedAdam and FedAdagrad [121].

1: Initialization: 𝑥0 , 𝑣−1 ≥ 𝜏 2 ,
2: Decay parameters 𝛽1 , 𝛽2 ∈ [0, 1]
3: for 𝑡 = 0, ...., 𝑇 − 1 do
4: 𝑆𝑡 : Sample subset of clients
5: 𝑥𝑡𝑖,0 = 𝑥𝑡
6: for each client 𝑖 ∈ 𝑆 in parallel do
7: for 𝑘 = 0, ..., 𝐾 − 1 do
𝑡
8: Compute an unbiased estimate 𝑔𝑖,𝑘 of ∇𝐹𝑖 (𝑥𝑡𝑖,𝑘 )
9: 𝑥𝑡𝑖,𝑘+1 = 𝑥𝑡𝑖,𝑘 - 𝜂𝑙 𝑔𝑖,𝑘
𝑡

10: Δ𝑡𝑖 = 𝑥𝑡𝑖,𝐾 -𝑥𝑡

∑
11:
1
Δ𝑡 = 𝛽1 Δ𝑡−1 + (1 − 𝛽1 )( ∣𝑆∣ 𝑖∈𝑆 Δ𝑡𝑖 )
12: 𝑣𝑡 = 𝑣𝑡−1 − (1 − 𝛽2 )Δ2𝑡 𝑠𝑖𝑔𝑛(𝑣𝑡−1 − Δ2𝑡 )(FedYogi)
13: 𝑣𝑡 = 𝛽2 𝑣𝑡−1 + (1 − 𝛽2 )Δ2𝑡 (FedAdam)
14: 𝑣𝑡 = 𝑣𝑡−1 + Δ2𝑡 (FedAdagrad)
Δ
15: 𝑥𝑡+1 = 𝑥𝑡 + 𝜂 √𝑣 𝑡+𝜏
𝑡
16: end for
17: end for
18: end for

20
B. Yurdem, M. Kuzlu, M.K. Gullu et al. Heliyon 10 (2024) e38137

Algorithm 3 FedProx [122].

1: Input Parameters:
2: The number of devices are denoted by K,
3: Total round is denoted as T,
4: 𝜇 is used as proximal term,
5: 𝛾 is used for step size,
6: 𝜔0 is donated for initial model,
7: N is donates for the total number of device,
8: 𝜌𝑘 is used as device-specific weights for K=1,....,N
9: for 𝑡 = 0, ...., 𝑇 − 1 do
10: A random subset is selected by subset 𝑆𝑡 of K
11: (each device 𝑘 is chosen with probability 𝑝𝑘 )
12: 𝜔𝑡 is send by server to all chosen devices
13: Each chosen device 𝑘 ∈ 𝑆𝑡 finds a 𝜔𝑡+1𝑘
which is a 𝛾𝑘𝑡 - inexact minimizer of:
𝜇
14: 𝜔𝑡+1
𝑘
≈ 𝑎𝑟𝑔𝑚𝑖𝑛𝜔 ℎ𝑘 (𝜔; 𝜔𝑡 ) = 𝐹𝑘 (𝜔) + 2
∣∣ 𝜔 − 𝜔𝑡 ∣∣2
15: 𝑘 ∈ 𝑆𝑡 is sent by each device 𝜔𝑡+1
𝑘
back to the server
16: Each server is aggregated by the 𝜔 as 𝜔𝑡+1
∑ 𝑡+1
17: 𝜔𝑡+1 = 𝐾 𝑘∈𝑆𝑡 𝜔𝑘
1

18: end for

Algorithm 4 Scaffold [123].

1: Initial Input Parameters:
2: Initial model is denoted by 𝑥𝑜
3: Scalar value is denoted by c
4: A local learning rate or step size is denoted by 𝜂𝜄
5: A global learning rate or step size is denoted by 𝜂
6: for 𝑡 = 0, ...., 𝑇 − 1 do
7: A subset S of clients is sampled
8: 𝑥𝑡𝑖 = 𝑥𝑡 = 𝑥𝑡 5 ⊳ Parallel processing over clients is done
9: for each client 𝑖 ∈ 𝑆 do in parallel
10: for 𝑒 = 1, ..., E do
11: for 𝑏 ∈ 𝐵𝑖 do
12: 𝑔𝑖𝑡 = ∇𝑓𝑖 (𝑥𝑡𝑖 ; 𝑏)
13: 𝑥𝑡𝑖 = 𝑥𝑡𝑖 − 𝜂𝜄(𝑔𝑖𝑡 − 𝑐𝑖 + 𝐶)
14: 𝑐𝑖+ = 𝑐𝑖 − 𝑐 + (𝐸 ∣ 𝐵𝑖 ∣ 𝜂𝜄)−1 (𝑥𝑡𝑖 − 𝑥𝑖 )
15: Δ𝑥𝑖 = 𝑥𝑡𝑖 − 𝑥𝑡 , 𝑐𝑖 = 𝑐𝑖+ − 𝑐𝑖
16: 𝑐𝑖 = 𝑐𝑖+
∑ ∑ 𝑛 ∑ 𝑛𝑖
17: 𝑛 = 𝑖∈𝑆 𝑛𝑖 , Δ𝑥 = 𝑖∈𝑆 𝑛𝑖 Δ𝑥𝑖 , Δ𝑐 = 𝑖∈𝑆 𝑛
Δ𝑐𝑖
∣𝑆∣
18: 𝑥𝑡+1 = 𝑥𝑡 + 𝜂Δ𝑥, 𝑐 = 𝑐 + 𝑁
Δ𝑐
19: end for
20: end for
21: end for
22: end for

References

[1] D. Silver, A. Huang, C.J. Maddison, A. Guez, L. Sifre, G. van den Driessche, J. Schrittwieser, I. Antonoglou, V. Panneershelvam, M. Lanctot, S. Dieleman, D.
Grewe, J. Nham, N. Kalchbrenner, I. Sutskever, T. Lillicrap, M. Leach, K. Kavukcuoglu, T. Graepel, D. Hassabis, Mastering the game of Go with deep neural
networks and tree search, Nature 529 (28) (2016) 484–492.
[2] L. Li, Y. Fan, M. Tse, K.-Y. Lin, A review of applications in federated learning, Comput. Ind. Eng. 149 (2020) 106854.
[3] P. Kairouz, H.B. McMahan, B. Avent, A. Bellet, M. Bennis, A.N. Bhagoji, K. Bonawitz, Z. Charles, G. Cormode, R. Cummings, et al., Advances and open problems
in federated learning, Found. Trends Mach. Learn. 14 (1–2) (2021) 1–210.
[4] J. Konečnỳ, H.B. McMahan, D. Ramage, P. Richtárik, Federated optimization: distributed machine learning for on-device intelligence, arXiv preprint, arXiv:
1610.02527, 2016.
[5] R. Shokri, V. Shmatikov, Privacy-preserving deep learning, in: Proceedings of the 22nd ACM SIGSAC Conference on Computer and Communications Security,
2015, pp. 1310–1321.
[6] Q. Yang, Y. Liu, T. Chen, Y. Tong, Federated machine learning: concept and applications, ACM Trans. Intell. Syst. Technol. 10 (2) (2019) 1–19.
[7] P. Regulation, Regulation (eu) 2016/679 of the European Parliament and of the council, Regulation (eu) 679 (2016) 2016.
[8] M. Boban, Digital single market and eu data protection reform with regard to the processing of personal data as the challenge of the modern world, in: Economic
and Social Development: Book of Proceedings, 2016, p. 191.
[9] C. Zhang, X. Hu, Y. Xie, M. Gong, B. Yu, A privacy-preserving multi-task learning framework for face detection, landmark localization, pose estimation, and
gender recognition, Front. Neurorobot. 13 (2020) 112.
[10] Y. Xie, H. Wang, B. Yu, C. Zhang, Secure collaborative few-shot learning, Knowl.-Based Syst. 203 (2020) 106157.
[11] W.G. Van Panhuis, P. Paul, C. Emerson, J. Grefenstette, R. Wilder, A.J. Herbst, D. Heymann, D.S. Burke, A systematic review of barriers to data sharing in public
health, BMC Public Health 14 (1) (2014) 1–9.
[12] S. Sarp, M. Kuzlu, E. Wilson, O. Guler, Wg2an: synthetic wound image generation using generative adversarial network, J. Eng. 2021 (5) (2021) 286–294.
[13] S. Sarp, Y. Zhao, M. Kuzlu, Artificial Intelligence-Powered Chronic Wound Management System: Towards Human Digital Twins, 2022.
[14] S.D. Holcomb, W.K. Porter, S.V. Ault, G. Mao, J. Wang, Overview on deepmind and its alphago zero ai, in: Proceedings of the 2018 International Conference
on Big Data and Education, 2018, pp. 67–71.

21
B. Yurdem, M. Kuzlu, M.K. Gullu et al. Heliyon 10 (2024) e38137

[15] A. Krizhevsky, I. Sutskever, G.E. Hinton, Imagenet classification with deep convolutional neural networks, Commun. ACM 60 (6) (2017) 84–90.
[16] G. Hinton, L. Deng, D. Yu, G.E. Dahl, A.-r. Mohamed, N. Jaitly, A. Senior, V. Vanhoucke, P. Nguyen, T.N. Sainath, et al., Deep neural networks for acoustic
modeling in speech recognition: the shared views of four research groups, IEEE Signal Process. Mag. 29 (6) (2012) 82–97.
[17] A. Hard, K. Rao, R. Mathews, S. Ramaswamy, F. Beaufays, S. Augenstein, H. Eichner, C. Kiddon, D. Ramage, Federated learning for mobile keyboard prediction,
arXiv preprint, arXiv:1811.03604, 2018.
[18] B. McMahan, E. Moore, D. Ramage, S. Hampson, B.A. y Arcas, Communication-efficient learning of deep networks from decentralized data, in: Artificial
Intelligence and Statistics, PMLR, 2017, pp. 1273–1282.
[19] M. Abadi, A. Chu, I. Goodfellow, H.B. McMahan, I. Mironov, K. Talwar, L. Zhang, Deep learning with differential privacy, in: Proceedings of the 2016 ACM
SIGSAC Conference on Computer and Communications Security, 2016, pp. 308–318.
[20] J. Wang, Z. Charles, Z. Xu, G. Joshi, H.B. McMahan, M. Al-Shedivat, G. Andrew, S. Avestimehr, K. Daly, D. Data, et al., A field guide to federated optimization,
arXiv preprint, arXiv:2107.06917, 2021.
[21] J. Qi, Q. Zhou, L. Lei, K. Zheng, Federated reinforcement learning: techniques, applications, and open challenges, arXiv preprint, arXiv:2108.11887, 2021.
[22] Q. Ho, J. Cipar, H. Cui, S. Lee, J.K. Kim, P.B. Gibbons, G.A. Gibson, G. Ganger, E.P. Xing, More effective distributed ml via a stale synchronous parallel parameter
server, Adv. Neural Inf. Process. Syst. 26 (2013).
[23] S. Lee, M.E. Lacy, M. Jankowich, A. Correa, W.-C. Wu, Association between obesity phenotypes of insulin resistance and risk of type 2 diabetes in African
Americans: the Jackson heart study, J. Clin. Transl. Endocrinol. 19 (2020) 100210.
[24] A. Gascón, P. Schoppmann, B. Balle, M. Raykova, J. Doerner, S. Zahur, D. Evans, Privacy-preserving distributed linear regression on high-dimensional data,
Cryptol. ePrint Arch. (2016).
[25] D. Gao, C. Ju, X. Wei, Y. Liu, T. Chen, Q. Yang, Hhhfl: hierarchical heterogeneous horizontal federated learning for electroencephalography, arXiv preprint,
arXiv:1909.05784, 2019.
[26] W. Huang, T. Li, D. Wang, S. Du, J. Zhang, T. Huang, Fairness and accuracy in horizontal federated learning, Inf. Sci. 589 (2022) 170–185.
[27] H. Zhu, H. Zhang, Y. Jin, From federated learning to federated neural architecture search: a survey, Complex Intell. Syst. 7 (2) (2021) 639–657.
[28] M. Kuzlu, Z. Xiao, M. Tabassum, F.O. Catak, A robust diabetes mellitus prediction system based on federated learning strategies, in: 2023 International Conference
on Intelligent Computing, Communication, Networking and Services (ICCNS), 2023, pp. 246–253.
[29] J.O.d. Terrail, S.-S. Ayed, E. Cyffers, F. Grimberg, C. He, R. Loeb, P. Mangold, T. Marchand, O. Marfoq, E. Mushtaq, et al., Flamby: datasets and benchmarks
for cross-silo federated learning in realistic healthcare settings, arXiv preprint, arXiv:2210.04620, 2022.
[30] T. Li, A.K. Sahu, M. Zaheer, M. Sanjabi, A. Talwalkar, V. Smith, Federated optimization in heterogeneous networks, Proc. Mach. Learn. Syst. 2 (2020) 429–450.
[31] S.P. Karimireddy, S. Kale, M. Mohri, S. Reddi, S. Stich, A.T. Suresh, Scaffold: stochastic controlled averaging for federated learning, in: International Conference
on Machine Learning, PMLR, 2020, pp. 5132–5143.
[32] L. Gao, H. Fu, L. Li, Y. Chen, M. Xu, C.-Z. Xu, Feddc: federated learning with non-iid data via local drift decoupling and correction, in: Proceedings of the
IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 10112–10121.
[33] H. Nguyen, H. Warrier, Y. Gupta, A novel approach for federated learning with non-iid data, in: 2022 9th International Conference on Soft Computing & Machine
Intelligence (ISCMI), IEEE, 2022, pp. 62–67.
[34] T.K. Dang, X. Lan, J. Weng, M. Feng, Federated learning for electronic health records, ACM Trans. Intell. Syst. Technol. 13 (5) (2022) 1–17.
[35] C. Dwork, Differential privacy, in: International Colloquium on Automata, Languages, and Programming, Springer, 2006, pp. 1–12.
[36] F. Bayatbabolghani, M. Blanton, Secure multi-party computation, in: Proceedings of the 2018 ACM SIGSAC Conference on Computer and Communications
Security, 2018, pp. 2157–2159.
[37] G. Szegedi, P. Kiss, T. Horváth, Evolutionary federated learning on eeg-data, in: ITAT, 2019, pp. 71–78.
[38] E. Choi, M.T. Bahadori, L. Song, W.F. Stewart, J. Sun, Gram: graph-based attention model for healthcare representation learning, in: Proceedings of the 23rd
ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2017, pp. 787–795.
[39] S.R. Pfohl, A.M. Dai, K. Heller, Federated and differentially private learning for electronic health records, arXiv preprint, arXiv:1911.05861, 2019.
[40] S. Silva, B.A. Gutman, E. Romero, P.M. Thompson, A. Altmann, M. Lorenzi, Federated learning in distributed medical databases: meta-analysis of large-scale
subcortical brain data, in: 2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019), IEEE, 2019, pp. 270–274.
[41] H. Ye, L. Liang, G.Y. Li, J. Kim, L. Lu, M. Wu, Machine learning for vehicular networks: recent advances and application examples, IEEE Veh. Technol. Mag.
13 (2) (2018) 94–101.
[42] D.C. Nguyen, M. Ding, P.N. Pathirana, A. Seneviratne, J. Li, H.V. Poor, Federated learning for Internet of things: a comprehensive survey, IEEE Commun. Surv.
Tutor. 23 (3) (2021) 1622–1658.
[43] A. Gharaibeh, M.A. Salahuddin, S.J. Hussini, A. Khreishah, I. Khalil, M. Guizani, A. Al-Fuqaha, Smart cities: a survey on data management, security, and enabling
technologies, IEEE Commun. Surv. Tutor. 19 (4) (2017) 2456–2501.
[44] M. Mohammadi, A. Al-Fuqaha, Enabling cognitive smart cities using big data and machine learning: approaches and challenges, IEEE Commun. Mag. 56 (2)
(2018) 94–101.
[45] W. Wang, C. Feng, B. Zhang, H. Gao, Environmental monitoring based on fog computing paradigm and Internet of things, IEEE Access 7 (2019) 127154–127165.
[46] M. Masera, E.F. Bompard, F. Profumo, N. Hadjsaid, Smart (electricity) grids for smart cities: assessing roles and societal impacts, Proc. IEEE 106 (4) (2018)
613–625.
[47] S. Pérez, J. Pérez, P. Arroba, R. Blanco, J.L. Ayala, J.M. Moya, Predictive gpu-based adas management in energy-conscious smart cities, in: 2019 Ieee International
Smart Cities Conference (isc2), IEEE, 2019, pp. 349–354.
[48] N. Victor, M. Alazab, S. Bhattacharya, S. Magnusson, P.K.R. Maddikunta, K. Ramana, T.R. Gadekallu, et al., Federated learning for iout: concepts, applications,
challenges and opportunities, arXiv preprint, arXiv:2207.13976, 2022.
[49] H. Zhao, F. Ji, Q. Guan, Q. Li, S. Wang, H. Dong, M. Wen, Federated meta learning enhanced acoustic radio cooperative framework for ocean of things underwater
acoustic communications, arXiv preprint, arXiv:2105.13296, 2021.
[50] C.L. Hagerty, Deepwater Horizon Oil Spill: Selected Issues for Congress, Diane Publishing, 2010.
[51] J. Telford, J. Cosgrave, The international humanitarian system and the 2004 Indian Ocean earthquake and tsunamis, Disasters 31 (1) (2007) 1–28.
[52] M. Karagoz, Maritime security operations and the combined joint operations from the sea center of excellence, in: Maritime Security and Defence Against
Terrorism, IOS Press, 2012, pp. 87–91.
[53] W. Saad, M. Bennis, M. Chen, A vision of 6g wireless systems: applications, trends, technologies, and open research problems, IEEE Netw. 34 (3) (2019) 134–142.
[54] X. Han, H. Yu, H. Gu, Visual inspection with federated learning, in: Image Analysis and Recognition: 16th International Conference, ICIAR 2019, Waterloo, ON,
Canada, August 27–29, 2019, Proceedings, Part II 16, Springer, 2019, pp. 52–64.
[55] F. Liu, X. Wu, S. Ge, W. Fan, Y. Zou, Federated learning for vision-and-language grounding problems, in: Proceedings of the AAAI Conference on Artificial
Intelligence, vol. 34, 2020, pp. 11572–11579.
[56] N.I. Mowla, N.H. Tran, I. Doh, K. Chae, Federated learning-based cognitive detection of jamming attack in flying ad-hoc network, IEEE Access 8 (2019)
4338–4350.
[57] Y.M. Saputra, D.T. Hoang, D.N. Nguyen, E. Dutkiewicz, M.D. Mueck, S. Srikanteswara, Energy demand prediction with federated learning for electric vehicle
networks, in: 2019 IEEE Global Communications Conference (GLOBECOM), IEEE, 2019, pp. 1–6.
[58] W. Yang, Y. Zhang, K. Ye, L. Li, C.-Z. Xu, Ffd: a federated learning based method for credit card fraud detection, in: Big Data–BigData 2019: 8th International
Congress, Held as Part of the Services Conference Federation, SCF 2019, San Diego, CA, USA, June 25–30, 2019, Proceedings 8, Springer, 2019, pp. 18–32.

22
B. Yurdem, M. Kuzlu, M.K. Gullu et al. Heliyon 10 (2024) e38137

[59] M.R. Sprague, A. Jalalirad, M. Scavuzzo, C. Capota, M. Neun, L. Do, M. Kopp, Asynchronous federated learning for geospatial applications, in: Joint European
Conference on Machine Learning and Knowledge Discovery in Databases, Springer, 2018, pp. 21–28.
[60] Y. Qu, M.P. Uddin, C. Gan, Y. Xiang, L. Gao, J. Yearwood, Blockchain-enabled federated learning: a survey, ACM Comput. Surv. 55 (4) (2022) 1–35.
[61] S. Singh, S. Rathore, O. Alfarraj, A. Tolba, B. Yoon, A framework for privacy-preservation of iot healthcare data using federated learning and blockchain
technology, Future Gener. Comput. Syst. 129 (2022) 380–388.
[62] C. Ma, J. Li, L. Shi, M. Ding, T. Wang, Z. Han, H.V. Poor, When federated learning meets blockchain: a new distributed learning paradigm, IEEE Comput. Intell.
Mag. 17 (3) (2022) 26–33.
[63] D. Hou, J. Zhang, K.L. Man, J. Ma, Z. Peng, A systematic literature review of blockchain-based federated learning: architectures, applications and issues, in:
2021 2nd Information Communication Technologies Conference (ICTC), IEEE, 2021, pp. 302–307.
[64] H.G. Abreha, M. Hayajneh, M.A. Serhani, Federated learning in edge computing: a systematic survey, Sensors 22 (2) (2022) 450.
[65] J. Zhang, B. Chen, X. Cheng, H.T.T. Binh, S. Yu, Poisongan: generative poisoning attacks against federated learning in edge computing systems, IEEE Int. Things
J. 8 (5) (2020) 3310–3322.
[66] Y. Ye, S. Li, F. Liu, Y. Tang, W. Hu, Edgefed: optimized federated learning based on edge computing, IEEE Access 8 (2020) 209191–209198, https://doi.org/
10.1109/ACCESS.2020.3038287.
[67] D. Ye, R. Yu, M. Pan, Z. Han, Federated learning in vehicular edge computing: a selective model aggregation approach, IEEE Access 8 (2020) 23920–23935,
https://doi.org/10.1109/ACCESS.2020.2968399.
[68] Tensorflow federated: machine learning on decentralized data, https://www.tensorflow.org/federated, 2020. (Accessed 21 May 2023).
[69] A. Ziller, A. Trask, A. Lopardo, B. Szymkow, B. Wagner, E. Bluemke, J.-M. Nounahon, J. Passerat-Palmbach, K. Prakash, N. Rose, et al., Pysyft: a library for easy
federated learning, in: Federated Learning Systems: Towards Next-Generation AI, 2021, pp. 111–139.
[70] H.R. Roth, Y. Cheng, Y. Wen, I. Yang, Z. Xu, Y.-T. Hsieh, K. Kersten, A. Harouni, C. Zhao, K. Lu, et al., Nvidia flare: federated learning from simulation to
real-world, arXiv preprint, arXiv:2210.13291, 2022.
[71] Y. Liu, T. Fan, T. Chen, Q. Xu, Q. Yang, Fate: an industrial grade platform for collaborative learning with data protection, J. Mach. Learn. Res. 22 (1) (2021)
10320–10325.
[72] D.J. Beutel, T. Topal, A. Mathur, X. Qiu, J. Fernandez-Marques, Y. Gao, L. Sani, K.H. Li, T. Parcollet, P.P.B. de Gusmão, et al., Flower: a friendly federated
learning research framework, arXiv preprint, arXiv:2007.14390, 2020.
[73] H. Ludwig, N. Baracaldo, G. Thomas, Y. Zhou, A. Anwar, S. Rajamoni, Y. Ong, J. Radhakrishnan, A. Verma, M. Sinn, et al., Ibm federated learning: an enterprise
framework white paper v0. 1, arXiv preprint, arXiv:2007.10987, 2020.
[74] D. Zeng, S. Liang, X. Hu, H. Wang, Z. Xu, Fedlab: a flexible federated learning framework, J. Mach. Learn. Res. 24 (100) (2023) 1–7.
[75] C. He, S. Li, J. So, X. Zeng, M. Zhang, H. Wang, X. Wang, P. Vepakomma, A. Singh, H. Qiu, et al., Fedml: a research library and benchmark for federated
machine learning, arXiv preprint, arXiv:2007.13518, 2020.
[76] M. Hipolito Garcia, A. Manoel, D. Madrigal Diaz, F. Mireshghallah, R. Sim, D. Dimitriadis, Flute: a scalable, extensible framework for high-performance federated
learning simulations, arXiv e-prints (2022) arXiv–2203.
[77] G.A. Reina, A. Gruzdev, P. Foley, O. Perepelkina, M. Sharma, I. Davidyuk, I. Trushkin, M. Radionov, A. Mokrov, D. Agapov, et al., Openfl: an open-source
framework for federated learning, arXiv preprint, arXiv:2105.06413, 2021.
[78] K. Cheng, T. Fan, Y. Jin, Y. Liu, T. Chen, D. Papadopoulos, Q. Yang, Secureboost: a lossless federated learning framework, IEEE Intell. Syst. 36 (6) (2021) 87–98.
[79] J. Benet, Ipfs-content addressed, versioned, p2p file system, arXiv preprint, arXiv:1407.3561, 2014.
[80] S. Silva, A. Altmann, B. Gutman, M. Lorenzi, Fed-biomed: a general open-source frontend framework for federated learning in healthcare, in: Domain Adaptation
and Representation Transfer, and Distributed and Collaborative Learning: Second MICCAI Workshop, DART 2020, and First MICCAI Workshop, DCL 2020, Held
in Conjunction with MICCAI 2020, Lima, Peru, October 4–8, 2020, Proceedings 2, Springer, 2020, pp. 201–210.
[81] Y. Xie, Z. Wang, D. Gao, D. Chen, L. Yao, W. Kuang, Y. Li, B. Ding, J. Zhou, Federatedscope: a flexible federated learning platform for heterogeneity, arXiv
preprint, arXiv:2204.05011, 2022.
[82] I. Kholod, E. Yanaki, D. Fomichev, E. Shalugin, E. Novikova, E. Filippov, M. Nordlund, Open-source federated learning frameworks for iot: a comparative review
and analysis, Sensors 21 (1) (2020) 167.
[83] X. Liu, T. Shi, C. Xie, Q. Li, K. Hu, H. Kim, X. Xu, B. Li, D. Song, Unifed: a benchmark for federated learning frameworks, arXiv preprint, arXiv:2207.10308,
2022.
[84] Z. Zuo, M. Watson, D. Budgen, R. Hall, C. Kennelly, N. Al Moubayed, Data anonymization for pervasive health care: systematic literature mapping study, JMIR
Med. Inform. 9 (10) (2021) e29871.
[85] NVIDIA, NVFlare: a framework for federated learning, https://developer.nvidia.com/nvflare, 2021. (Accessed 6 July 2023).
[86] Experimenting with novel distributed applications using NVIDIA Flare 2.1, https://developer.nvidia.com/blog/experimenting-with-novel-distributed-
applications-using-nvidia-flare-2-1/, 2022. (Accessed 22 August 2023).
[87] F. Zheng, K. Li, J. Tian, X. Xiang, et al., A vertical federated learning method for interpretable scorecard and its application in credit scoring, arXiv preprint,
arXiv:2009.06218, 2020.
[88] Z. Xiong, Z. Cheng, C. Xu, X. Lin, X. Liu, D. Wang, X. Luo, Y. Zhang, N. Qiao, M. Zheng, et al., Facing small and biased data dilemma in drug discovery with
federated learning, BioRxiv (2020) 2020–03.
[89] Architecture, https://fate.readthedocs.io/en/latest/architecture/, 2021. (Accessed 22 August 2023).
[90] Flower architecture, https://flower.dev/docs/framework/contributor-explanation-architecture.html, 2022. (Accessed 22 August 2023).
[91] Federated learning architecture, https://dataplatform.cloud.ibm.com/docs/content/wsj/analyze-data/fl-arch.html, 2023. (Accessed 22 August 2023).
[92] FLUTE overview, https://microsoft.github.io/msrflute/overview, 2021. (Accessed 22 August 2023).
[93] W. Riviera, I.B. Galazzo, G. Menegaz, Felebrities: a user-centric assessment of federated learning frameworks, IEEE Access (2023).
[94] C. Zhang, Y. Xie, H. Bai, B. Yu, W. Li, Y. Gao, A survey on federated learning, Knowl.-Based Syst. 216 (2021) 106775.
[95] Y. Aono, T. Hayashi, L. Wang, S. Moriai, et al., Privacy-preserving deep learning via additively homomorphic encryption, IEEE Trans. Inf. Forensics Secur.
13 (5) (2017) 1333–1345.
[96] K. Bonawitz, V. Ivanov, B. Kreuter, A. Marcedone, H.B. McMahan, S. Patel, D. Ramage, A. Segal, K. Seth, Practical secure aggregation for privacy-preserving
machine learning, in: Proceedings of the 2017 ACM SIGSAC Conference on Computer and Communications Security, 2017, pp. 1175–1191.
[97] T. Nguyen, M.T. Thai, Preserving privacy and security in federated learning, IEEE/ACM Trans. Netw. (2023).
[98] K. Bonawitz, V. Ivanov, B. Kreuter, A. Marcedone, H.B. McMahan, S. Patel, D. Ramage, A. Segal, K. Seth, Practical secure aggregation for federated learning on
user-held data, arXiv preprint, arXiv:1611.04482, 2016.
[99] Z. Wang, M. Song, Z. Zhang, Y. Song, Q. Wang, H. Qi, Beyond inferring class representatives: user-level privacy leakage from federated learning, in: IEEE
INFOCOM 2019-IEEE Conference on Computer Communications, IEEE, 2019, pp. 2512–2520.
[100] R.C. Geyer, T. Klein, M. Nabi, Differentially private federated learning: a client level perspective, arXiv preprint, arXiv:1712.07557, 2017.
[101] S. Niknam, H.S. Dhillon, J.H. Reed, Federated learning for wireless communications: motivation, opportunities, and challenges, IEEE Commun. Mag. 58 (6)
(2020) 46–51.
[102] K. Hsieh, A. Harlap, N. Vijaykumar, D. Konomis, G.R. Ganger, P.B. Gibbons, O. Mutlu, Gaia: geo-distributed machine learning approaching LAN speeds, in: 14th
USENIX Symposium on Networked Systems Design and Implementation (NSDI 17), 2017, pp. 629–647.

23
B. Yurdem, M. Kuzlu, M.K. Gullu et al. Heliyon 10 (2024) e38137

[103] W. Luping, W. Wei, L. Bo, Cmfl: mitigating communication overhead for federated learning, in: 2019 IEEE 39th International Conference on Distributed
Computing Systems (ICDCS), IEEE, 2019, pp. 954–964.
[104] P.M. Mammen, Federated learning: opportunities and challenges, arXiv preprint, arXiv:2101.05428, 2021.
[105] G. Sun, Y. Cong, J. Dong, Q. Wang, L. Lyu, J. Liu, Data poisoning attacks on federated machine learning, IEEE Int. Things J. 9 (13) (2021) 11365–11375.
[106] M. Fang, X. Cao, J. Jia, N. Gong, Local model poisoning attacks to Byzantine-robust federated learning, in: 29th USENIX Security Symposium (USENIX Security
20), 2020, pp. 1605–1622.
[107] E. Bagdasaryan, A. Veit, Y. Hua, D. Estrin, V. Shmatikov, How to backdoor federated learning, in: International Conference on Artificial Intelligence and
Statistics, PMLR, 2020, pp. 2938–2948.
[108] X. Zhou, M. Xu, Y. Wu, N. Zheng, Deep model poisoning attack on federated learning, Future Internet 13 (3) (2021) 73.
[109] A.N. Bhagoji, S. Chakraborty, P. Mittal, S. Calo, Analyzing federated learning through an adversarial lens, in: International Conference on Machine Learning,
PMLR, 2019, pp. 634–643.
[110] K. Bonawitz, H. Eichner, W. Grieskamp, D. Huba, A. Ingerman, V. Ivanov, C. Kiddon, J. Konečnỳ, S. Mazzocchi, B. McMahan, et al., Towards federated learning
at scale: system design, Proc. Mach. Learn. Syst. 1 (2019) 374–388.
[111] Y. Liu, J. James, J. Kang, D. Niyato, S. Zhang, Privacy-preserving traffic flow prediction: a federated learning approach, IEEE Int. Things J. 7 (8) (2020)
7751–7763.
[112] H. Yu, Z. Liu, Y. Liu, T. Chen, M. Cong, X. Weng, D. Niyato, Q. Yang, A sustainable incentive scheme for federated learning, IEEE Intell. Syst. 35 (4) (2020)
58–69.
[113] J. Kang, Z. Xiong, D. Niyato, H. Yu, Y.-C. Liang, D.I. Kim, Incentive design for efficient federated learning in mobile networks: a contract theory approach, in:
2019 IEEE VTS Asia Pacific Wireless Communications Symposium (APWCS), IEEE, 2019, pp. 1–5.
[114] N. Kourtellis, K. Katevas, D. Perino, Flaas: federated learning as a service, in: Proceedings of the 1st Workshop on Distributed Machine Learning, 2020, pp. 7–13.
[115] C. Xie, S. Koyejo, I. Gupta, Asynchronous federated optimization, arXiv preprint, arXiv:1903.03934, 2019.
[116] H.B. Desai, M.S. Ozdayi, M. Kantarcioglu, Blockfla: accountable federated learning via hybrid blockchain architecture, in: Proceedings of the Eleventh ACM
Conference on Data and Application Security and Privacy, 2021, pp. 101–112.
[117] C. Zhu, J. Zhang, X. Sun, B. Chen, W. Meng, Adfl: defending backdoor attacks in federated learning via adversarial distillation, Comput. Secur. 132 (2023)
103366.
[118] J. Zhang, J. Chen, D. Wu, B. Chen, S. Yu, Poisoning attack in federated learning using generative adversarial nets, in: 2019 18th IEEE International Conference on
Trust, Security and Privacy in Computing and Communications/13th IEEE International Conference on Big Data Science and Engineering (TrustCom/BigDataSE),
2019, pp. 374–380.
[119] M. Fang, X. Cao, J. Jia, N.Z. Gong, Local model poisoning attacks to byzantine-robust federated learning, CoRR, arXiv:1911.11815, 2019.
[120] B. Luo, Y. Liu, L. Wei, Q. Xu, Towards imperceptible and robust adversarial example attacks against neural networks, in: Proceedings of the AAAI Conference
on Artificial Intelligence, vol. 32, 2018.
[121] S. Reddi, Z. Charles, M. Zaheer, Z. Garrett, K. Rush, J. Konečnỳ, S. Kumar, H.B. McMahan, Adaptive federated optimization, arXiv preprint, arXiv:2003.00295,
2020.
[122] Should the global model replace the client model? ⋅ Issue #25 ⋅ litian96/FedProx — github.com, https://github.com/litian96/FedProx/issues/25. (Accessed 2
April 2023).
[123] S. Praneeth Karimireddy, S. Kale, M. Mohri, S.J. Reddi, S.U. Stich, A. Theertha Suresh, Scaffold: stochastic controlled averaging for federated learning, arXiv
e-prints (2019) arXiv–1910.

24