Open Access Library Journal

2024, Volume 11, e12239

ISSN Online: 2333-9721
ISSN Print: 2333-9705

Advances, Challenges & Recent Developments

in Federated Learning

Nsie Erimola María Reina Agripina*, Hua Shen, Blessed Shinga Mafukidze

Department of Computer Science, Hubei University of Technology, Wuhan, China

How to cite this paper: Agripina, N.E.M.R., Abstract

Shen, H. and Mafukidze, B.S. (2024) Advan-
ces, Challenges & Recent Developments in This has led to the rise of a paradigm shift in machine learning called federated
Federated Learning. Open Access Library learning (FL) that allows for decentralized model training over distributed data
Journal, 11: e12239. sources. With FL, devices, servers, or edges train the model together without
https://doi.org/10.4236/oalib.1112239
sharing their privacy-sensitive data, effectively addressing the arising data pri-
Received: September 4, 2024 vacy regulation, data residency, and data silos types of issues, among many
Accepted: October 18, 2024 others. The FL ecosystem has also been through a series of significant develop-
Published: October 21, 2024
ments, leading to the emergence of secure aggregation protocols and federated
Copyright © 2024 by author(s) and Open
optimization techniques for better model convergence and performance,
Access Library Inc. though there are still critical roadblocks such as data heterogeneity, communi-
This work is licensed under the Creative cation overhead, and vulnerability to attacks. This paper aims to summarize
Commons Attribution International
the current progress, practical limitations, and future research directions on
License (CC BY 4.0).
http://creativecommons.org/licenses/by/4.0/ the application of FL, particularly in the healthcare, finance, and Internet of
Open Access Things domains, as a means of preserving privacy and enhancing learning. The
future entails the incorporation of edge computing, decentralized learning
frameworks, and privacy-preserving techniques into the picture that has the
potential to reshape today’s state-of-the-art FL.

Subject Areas
Information Management, Machine Learning

Keywords
Federated Learning, Decentralized Technology, Machine Learning, Data

1. Introduction
Federated learning has emerged as a transformative paradigm in machine learn-
ing, offering a decentralized approach to collaboratively train models across dis-
tributed data sources while preserving data privacy and security. This innovative

DOI: 10.4236/oalib.1112239 Oct. 21, 2024 1 Open Access Library Journal

 N. E. M. R. Agripina et al.

framework enables organizations to leverage data from multiple devices, servers,

or edge devices without centralizing sensitive information, addressing the chal-
lenges associated with data silos, privacy regulations, and data residency require-
ments. Recent interest in this field has been high, both from research and an ap-
plied perspective. This monograph outlines the distinctive features and difficulties
of the federated learning environment, emphasizes significant practical limita-
tions and considerations, and then lists a variety of fruitful research directions.
This work’s objectives are to draw attention to research issues that are both theo-
retically and practically significant and to encourage research on problems that
could have significant real-world impact.
The International Data Corporation predicts that by 2025, there will be 80 bil-
lion Internet-connected devices, and the amount of data generated could reach
180 trillion gigabytes. On the other hand, those computation-intensive machine
learning applications can finally be run on devices rather than centralized cloud
data centers thanks to the evolution of powerful computation hardware design
and efficient computing architecture, like parallel high-performance graphics pro-
cessing units (GPUs), which also promotes the widespread use of machine learn-
ing. However, it is known that a machine learning model’s training requires a sig-
nificant amount of data, whereas the amount of data produced by a single device
is constrained. In addition, because of the diverse property. The machine learning
model created from one device is difficult to work with and desirable for others in
terms of individual behavior characteristics.

1.1. Overview of Federated Learning (FL)

Federated Learning (FL) has emerged as a promising approach for collaborative
model training across distributed devices while preserving data privacy and secu-
rity. This field has witnessed significant advances, notable challenges, and ongoing
developments that are shaping the future of decentralized machine learning. One
of the key advances in FL research is the development of more efficient algorithms
and protocols for secure aggregation of model updates from multiple clients. Se-
cure aggregation techniques, such as encryption and differential privacy, help en-
sure that individual contributions remain private while still enabling collaborative
model learning. Additionally, advancements in federated optimization methods,
like Federated Averaging and Federated Learning with Bayesian Neural Networks,
have improved model convergence and performance across distributed datasets
McMahan (2016) [1].
Despite these advances, FL faces several challenges that need to be addressed to
realize its full potential. One of the primary challenges is the heterogeneity of data
distributions and quality across devices, which can lead to biases and model per-
formance disparities. Research efforts are ongoing to develop robust federated
learning frameworks that can mitigate the impact of data distribution mismatches
and improve model generalization across diverse data sources. FL systems must
also contend with issues related to communication and computational overhead,

DOI: 10.4236/oalib.1112239 2 Open Access Library Journal

 N. E. M. R. Agripina et al.

as well as network failures and adversarial attacks that could compromise model
privacy and integrity. To address these challenges, researchers are exploring novel
solutions, such as decentralized learning scheduling algorithms, adaptive feder-
ated optimizers, and secure federated learning frameworks with differential pri-
vacy guarantees He (2020) [2]. Recent developments in FL research have focused
on expanding the applicability of federated learning to new domains, such as
healthcare, financial services, and IoT networks, where data privacy and regula-
tory compliance are paramount. Collaborative efforts between academia, indus-
try, and policymakers are driving the development of standards and best practices
for federated learning deployment in real-world settings. Additionally, the inte-
gration of edge computing capabilities and federated learning technologies is pav-
ing the way for decentralized AI applications that deliver personalized services
and insights at the network edge Yang (2019) [3].
Federated learning continues to evolve as a transformative paradigm for collab-
orative and privacy-preserving machine learning. Advances in algorithms, chal-
lenges in data heterogeneity and security, and recent developments in new appli-
cation domains are shaping the landscape of FL research and innovation, laying
the foundation for the next generation of decentralized AI systems.

1.2. Exploring the Evolution of Edge Computing: A Focus on

Federated Learning (FL)
Edge computing is not a brand-new idea. In fact, basic query processing across
distributed, low-power devices has been studied for decades in the contexts of
computing at the edge, fog computing, and query processing in sensor networks.
In recent works, the idea of centrally training machine learning models while serv-
ing and storing them locally has also been explored. However, it is possible to use
improved local resources on each device as the storage and computational capac-
ities of distributed networks’ devices increase. Due to this, federated learning,
which investigates training statistical models on distant devices directly, is gaining
popularity.
In order to dispatch data distribution and computing resources effectively and
create an effective cooperation model, parameter servers, atypical distributed and
centralized technology, primarily require a central server Ho (2013) [4]. There is
a double communication overhead as a result of this type of centralized data pro-
cessing technology.
The data must be copied if they are obtained for training from many databases
that are distributed throughout the central server at first. Then, for distributed
computation, a central server will distribute data to each distributed client. It in-
tensifies the system’s rigorous assessments for computing power, storage, and
bandwidth. For cases in FL, each client is entirely independent, data is not allo-
cated by the center and the training process is not directed by the server. There-
fore, FL is an integrated technology that combines machine learning models with
data fusion through decentralized collaboration.

DOI: 10.4236/oalib.1112239 3 Open Access Library Journal

 N. E. M. R. Agripina et al.

2. Characteristics of Federating Learning (FL)

Universality for cross-organizational scenarios FL, as proposed by Google, is es-
sentially an encrypted distributed machine learning technique that enables users
to create a training model jointly while maintaining the underlying data locally.
The original then According to Yang (2019) [3], the idea of FL has been expanded
to include any privacy-preserving decentralized collaborative machine learning
techniques. As a result, FL can handle both vertically partitioned data according
to features and horizontally partitioned data according to samples in a collabora-
tive learning environment. Cross-organizational enterprises could be integrated
into the federal framework by expanding FL. For instance, a bank with infor-
mation on customers’ purchasing power could work with an online marketplace
that has information on product features to suggest purchases.

2.1. Decentralized Training

Federated Learning (FL) refers to the process of training machine learning models
on multiple devices or servers without the need to bring all the data to a central
location. Instead of sending raw data to a centralized server for model training,
FL distributes the learning process across individual devices or edge servers, where
data is stored and processed locally. In a decentralized training setup, each device
or server participates in model training by computing model updates based on its
local data and then sending only the updated model parameters to a central server
or aggregator. The central server aggregates these partial model updates from
multiple devices to create a global model that reflects insights from the diverse
data sources participating in the FL process. By keeping data decentralized and
localized on individual devices, decentralized training helps protect the privacy of
sensitive information. Personal data remains under the control of the user and is
not exposed to potential security risks associated with centralized data storage or
transmission. This approach also enables users to retain ownership and control
over their data while still benefiting from the collective intelligence captured in
the global model. This is illustrated in Figure 1.

Figure 1. Beltrán et al. (2023) [5], Decentralized Technology.

DOI: 10.4236/oalib.1112239 4 Open Access Library Journal

 N. E. M. R. Agripina et al.

Decentralized training in FL promotes collaboration and knowledge sharing

among participants without compromising data privacy, allowing organizations
to leverage distributed data sources for model improvement and performance en-
hancement. Additionally, this approach enhances scalability and efficiency by
overcoming challenges related to data silos, network latency, and data privacy reg-
ulations, making FL an attractive solution for privacy-conscious and data-sensi-
tive applications in various domains

2.2. Equality and Empowerment in Federated Learning for Shared

Prosperity
All parties benefit from equal standing and definite dominion in this framework
for cooperation in order to achieve shared prosperity. In terms of equality, the
person who has access to a large amount of data typically holds the dominating
position in collaborative training that is spread. Therefore, the preference for
companies with large amounts of data or photographs with various types of labels
could have a negative impact on the growth of collaborative learning in the indus-
trial field. Small and medium-sized businesses lack motivation for cooperative
training in deep learning networks because institutions with big data can manip-
ulate the prediction model. However, due to equality among all parties in FL, the
standing of these clients with modest data sets would be elevated.
In conclusion, as shown in Figure 1. FL is a decentralized solution that enables
dispersed customers or organizations to autonomously train a collaborative
model while maintaining localized data. This approach can assist business organ-
izations in sharing collaboration without transferring any raw data.

3. Classification of Federated Learning (FL)

3.1. Vertical Federated Learning
This method is employed when each device has a data set with unique attributes
derived from sample examples. For example, Vertical FL can be used to create a
shared machine learning model across two organizations that have data about the
same group of individuals but distinct feature sets. In healthcare, one entity may
have electronic health records containing patient demographics and medical his-
tory, while another entity may have lab test results. By using Vertical Federated
Learning, both entities can collaborate to train a model that predicts patient out-
comes or diagnoses without sharing their sensitive data.

3.2. Horizontal Federated Learning

Horizontal Federated Learning is a form of Federated Learning where multiple
clients or devices with similar data distributions collaborate to train a global ma-
chine learning model while keeping their data locally. In horizontal Federated
Learning, each client or device has its own data. This method is utilized when each
device has a data set with the same feature space but a separate collection of sam-
ple instances. This sort of learning, in which the participating mobile phones have

DOI: 10.4236/oalib.1112239 5 Open Access Library Journal

 N. E. M. R. Agripina et al.

various training data with the same properties, is used in the initial FL Google
keyboard use case. This is illustrated in Figure 2.

Figure 2. Shaheen et al. (2024) [6], Application of Federated learning: Taxonomy, Challenges, Research
Trends.

Horizontal Federated Learning is particularly useful when dealing with multi-

ple distributed data sources that share similar characteristics but cannot be com-
bined due to privacy or regulatory constraints.

3.2.1. Cross-Silo Federated Learning

When there are fewer participating devices and they are accessible throughout all
rounds, Cross-Silo Federated Learning is used. The training data may be in FL
format, either horizontally or vertically. Cross-silo is mostly utilized for scenarios
involving organizations. Works like making use of cross-silo to help FL create
their model. In Cross-Silo Federated Learning, each entity trains a local model on
its own data and then shares only model updates with other entities instead of raw
data. These model updates are aggregated and used to improve the global model,
which benefits from the diversity and richness of data from multiple sources with-
out compromising data privacy or security. For example, in the context of mobile
device manufacturers collaborating to improve predictive text suggestions, each
manufacturer may have data on the language usage patterns and typing habits of
their users. By utilizing Cross-Silo Federated Learning, these manufacturers can
collaboratively train a more accurate predictive text model by leveraging the
unique insights from each dataset without sharing raw user data.

3.2.2. Cross-Device Federated Learning

Scenarios with a large number of participating devices use Crossdevice Federated
Learning. Client selection and incentive designs are two prominent strategies re-
quired to support this kind of FL. In Cross-Device Federated Learning, the model

DOI: 10.4236/oalib.1112239 6 Open Access Library Journal

 N. E. M. R. Agripina et al.

is trained locally on each device using its respective data, and model updates are
then aggregated and shared among devices. This allows for personalized model
training while benefiting from the diversity of data from various devices. By shar-
ing only model updates and not raw data, privacy and security are maintained.
For example, consider a user who uses a smartphone, tablet, and laptop for various
tasks such as web browsing, email, and social media. These devices can collaborate
through Cross-Device Federated Learning to train a personalized, predictive text
model that adapts to the user’s typing patterns and language preferences across all
devices.

4. Practical Uses of Federated Learning

Federated Learning offers a versatile and privacy-preserving approach to collabo-
rative model training, making it suitable for a wide range of applications across
various industries where data privacy and security are paramount. Federated
learning is crucial for supporting distributed training data applications that deal
with privacy-sensitive data. Theoretically, it appears to be a great strategy for re-
solving issues with data stored on a single central server or location that arise in
typical machine learning models. This is illustrated in Figure 3.

Figure 3. Farooq et al. (2023) [7], an example application of federated learning for the task of next-word prediction
on mobile phones.

4.1. Practical Use in Autonomous Vehicle

Federated learning is used to create self-driving cars since it can make predictions
in real-time. The data may contain realtime updates on the state of the roads and
traffic, enabling continual learning and quicker decision-making. This might lead
to a safer and more fun self-driving car experience. A prospective field for the
application of federated machine learning is the automotive industry. But at the
moment, research is the only thing being done in this area. Federated learning
may reduce training time for predicting the steering angle of self-driving cars, ac-
cording to one study.

DOI: 10.4236/oalib.1112239 7 Open Access Library Journal

 N. E. M. R. Agripina et al.

4.2. Practical Use in Healthcare

For machine learning applications, electronic health records (EHR) are regarded
as the primary source of healthcare data. If only the scant data present in a partic-
ular hospital is used to train ML models. The predictions may become somewhat
biased as a result.
As a result, training with more data is necessary to make the models more gen-
eralizable. This can be accomplished by exchanging data among businesses. Shar-
ing patient electronic health information between hospitals may not be possible
due to the sensitive nature of healthcare data. Federated learning may be an alter-
native for creating a collaborative learning model for healthcare data in such cir-
cumstances. This method may leverage a significant quantity of data from multi-
ple healthcare databases and devices to construct AI models while remaining com-
pliant with rules.

4.3. Practical Use in Finance

In the banking industry, federated learning is used most effectively for loan risk
assessment. The majority of the time, banks employ whitelisting strategies to ex-
clude clients by using credit card records from the central banks. Elements like
taxation and reputation etc., can be used to manage risk by working with other
financial institutions and e-commerce businesses. Since it is risky for businesses
to exchange customers’ sensitive information, they can utilize FL to create a ma-
chine-learning model for risk assessment. It offers a secure and collaborative ap-
proach to data analysis in the finance industry, enabling financial institutions to
leverage insights from multiple sources while ensuring data privacy and confiden-
tiality. By embracing Federated Learning, financial organizations can drive inno-
vation, improve decision-making processes, and enhance customer experiences in
a secure and compliant manner.

5. Security and Privacy Concerns

While Federated Learning offers numerous benefits in terms of privacy-preserv-
ing collaborative model training, there are also several security and privacy con-
cerns that need to be addressed to ensure the safe deployment of this technology.
Federated Learning models are vulnerable to assaults, just like any machine learn-
ing model. A central server that has been compromised or local learning devices
that have been compromised can both introduce assaults by the FL workflow’s
participants or by the framework. This is illustrated in Figure 4.
1) Attacks from the Back
In federated learning, secure averaging enables device anonymity during model
updating. A device or a set of devices might introduce a backdoor by sharing the
same functionality capabilities in the federated learning approach globally. An at-
tacker can mislabel some jobs using a backdoor without impacting the correctness
of the overall model. An attacker could, for instance, select a particular label for a
data instance with a certain set of traits. This is illustrated in Figure 5.

DOI: 10.4236/oalib.1112239 8 Open Access Library Journal

 N. E. M. R. Agripina et al.

Figure 4. Mothukuri et al. (2021) [8] FL Security and Privacy.

Figure 5. Pasquini & Böhme (2020) [9] Illustration of backdoor attacks.

Although the aforementioned algorithms could prevent an adversary from in-

vading a central server or clients, the encrypted parameters could still result in
information leakage due to new attack techniques. Many hybrid strategies to im-
prove the framework’s privacy have been put forth. However, accuracy may suffer
due to the noise that differential privacy introduces. The Hybrid-One approach uses
DP and MPC together to reduce noise without sacrificing accuracy, protecting

DOI: 10.4236/oalib.1112239 9 Open Access Library Journal

 N. E. M. R. Agripina et al.

communication messages by relying on MPC to do so, which introduces less noise

than typical local Truex (2019) [10]. As homomorphic encryption can be, this ap-
proach frequently has high communication costs and a slow convergence rate.
Sketched algorithms are also naturally suited for FL because original data must be
traced back through additional techniques and data IDs are not maintained. In
order to strengthen confidentially, Smith and Sekar (2019) [11] built a link be-
tween FL and the sketching algorithm. Targeted attacks also go by the name in
backdoor attacks. The severity of such attacks depends on the number of compro-
mised devices that are present and the federated learning model capabilities.
2) Data Poisoning Attack
Attempting to include some contaminated data, such as malicious samples or
data that has been camouflaged, in order to sabotage data integrity or cause bias
in training results. There are two primary assault modes known as “data poison-
ing,” including model skew and weaponization of feedback. Since an opponent
might directly modify the triggers to lead the global model astray, traditional ma-
chine learning systems are susceptible to data poisoning. However, as bad attack-
ers do not have direct access to raw data in FL, these conventional data poisoning
techniques are less effective or may require a large number of malevolent players
Smith, G. & Sekar (2019) [12].
3) Model Poisoning attack
Similar to data poisoning attacks, model poisoning attacks aim to contaminate
local models rather than local data. The model poisoning’s primary justification
attack involves introducing mistakes into the overall model. By compromising
some of the devices and altering its local model parameters, the adversary con-
ducts a model poisoning attack that reduces the accuracy of the global model.
4) Inferring Attack
This kind of attack’s key advantage is in its ability to recover training data
through a white box or a black box while also detecting privacy records. It can be
divided into two categories: reconstruction attacks and tracing attacks (sometimes
called membership inference attacks). The first of the two signals is whether to
assume that a client is present in the data set. The latter camp claims to have re-
trieved some participant-specific characteristics. Nasr (2019) [13] developed a
white-box membership inference attack approach that targets neural networks di-
rectly by utilizing the SGD vulnerability. The method was then successfully used
in a federated setup to infer information from an inquisitive server or any partic-
ipant.

6. Protection Method for FL Framework

Although the aforementioned algorithms could prevent an adversary from invad-
ing a central server or clients, the encrypted parameters could still result in infor-
mation leakage due to new attack techniques. Many hybrid strategies to improve
the framework’s privacy have been put forth. However, accuracy may suffer due
to the noise that differential privacy introduces. As homomorphic encryption can

DOI: 10.4236/oalib.1112239 10 Open Access Library Journal

 N. E. M. R. Agripina et al.

be, this approach frequently has high communication costs and a slow conver-
gence rate. Sketched algorithms are also naturally suited for FL because original
data must be traced back through additional techniques and data IDs are not
maintained in order to strengthen confidentially, Smith and Sekar (2019) [12]
built a link between FL and the sketching algorithm.

6.1. Recent Developments in FL

1) Incentive Mechanism
Current FL methodologies function under the presumption that gadgets will
assist in the learning process whenever necessary, regardless of the incentives. In
contrast, in actual practice, tools or Clients must receive financial compensation
for taking part. Works suggest a reputation-based incentive mechanism, in which
devices receive rewards based on their model accuracy, data reliability, and con-
tribution to the global model, to promote/improve device participation in FL.
These publications, however, omitted to discuss how to account for convergence
and new communication overheads incorporated into the architecture.
2) Blockchain in FL
An aggregator is required to update the global model and control the asynchro-
nous arrival of parameters from the devices. This may limit the FL models’ ability
to be widely adopted. As Blockchain is a decentralized network, without a central
aggregator, devices can work together to learn. Federated Learning is proposed
for works like those within a blockchain system.

6.2. Verification for Returned Model

Most privacy-preserving techniques in FL are predicated on the strong supposi-
tion that clients are semi-honest and follow instructions while maintaining an in-
terest in the collection of private data. Realistic application, however, obtains the
other kind. Clients may purposefully or unintentionally communicate an incor-
rect model, forcing the global model to vary from the expected trace. For instance,
in a wearable medical system, opponents could produce data that is convincing
but inaccurate in an effort to undermine the entire design Cai and Venkatasubra-
manian (2018) [14]. In FL, this particular Byzantine issue always arises. Therefore,
a Byzantine fault-tolerant system should be created, so that collaborative training
can continue to function even if certain clients don’t adhere to training protocol.

6.3. Federated Learning as a Service

These days, machine learning as a service is becoming more and more common,
and the majority of them only provide centralized services. It should take into
account application collaboration with third-party services in order to deliver
Federated Learning as a cloud service. Recent research attempted to develop an
FL framework (as a service) that enables contributions and teamwork on an ML
model from other parties. It is asserted that the framework can be used in any
operational setting. Some Key features and benefits of Federated Learning as a

DOI: 10.4236/oalib.1112239 11 Open Access Library Journal

 N. E. M. R. Agripina et al.

Service include Scalability, Cost-Effectiveness, Data Security and Privacy and

Model Aggregation and Management.

6.4. Feasibility and Application Potential of Federated Learning

Federated learning has shown high potential for practical applications, such as in
the areas of healthcare, finance, and the Internet of Things. The reasons it is likely
to be feasible in many areas are as follows: the FL paradigm is decentralized, so
the FL process is done locally on the devices of each participant. Federated learn-
ing involves privacy-preserving mechanisms and focuses on data from a substan-
tial number of users. In addition, FL is also implementable in a wide range of dis-
tributed platforms because of its scalability. Specifically, the following challenges
and opportunities are addressed every time in practical applications: data privacy
problems, such as utilizing decentralized data; security challenges inherent in FL’s
privacy-preserving mechanisms; computational efficiency challenges caused by
managing massive data; and communication overhead.
FLs effectiveness is greatly enhanced by its emphasis, on safeguarding data pri-
vacy and security as a priority factor in its viability as a technology solution for
industries like healthcare and finance that place value on protecting sensitive in-
formation. Unlike methods that involve sending data to a central repository for
analysis and model training which could compromise data security and privacy
concerns; FL ensures that data remains decentralized and secure, by sharing only
encrypted model updates. This functionality assists companies in adhering to pri-
vacy rules such, as GDPR and HIPAA which makes FL a practical choice, for busi-
nesses and establishments dealing with data security and confidentiality, which
are maintained in the implementation of FL allowing the development of models.
Dealing with differences in data is another obstacle that Federated Learning ef-
fectively deals with, which also makes it more feasible in practice. Federated
Learning is well suited for scenarios like networks of the Internet of Things (IoT)
or mobile devices where the data from each device can differ in terms of quality
or format. Traditional machine learning methods face challenges when handling
data that is not identically distributed (non IID). On the contrary, Federated
Learning frameworks are specifically crafted to adapt to these variations making
it possible to train models accurately and robustly across a range of data sources.
FL becomes an asset, for scenarios that involve cooperation among groups or de-
vices in diverse environments were maintaining consistency, in data gathering
and storage proves challenging.
Communication & computing well are super important for using FL in real life.
This is especially true in places like IoT networks where resources are tight. FL
helps by cutting down on the need to gather all data in one spot. But there’s a
catch: devices have to send updates about the model pretty often.
To make it work better, people have been working hard on ways to cut down
communication clutter, find the best times for updates, & make data smaller when
it’s sent between devices. Because of these upgrades, FL has become more resource-

DOI: 10.4236/oalib.1112239 12 Open Access Library Journal

 N. E. M. R. Agripina et al.

friendly and can grow easily. Now, it can be used in many different situations
where bandwidth & computing strength are low. This includes things like smart
gadgets, mobile apps, or industrial systems that are spread out.
In healthcare, federated learning (FL) holds great promise. It encourages team-
work among different institutions without putting patient privacy at risk. For in-
stance, think about several hospitals or medical centers teaming up to create
predictive models for diagnosing diseases or suggesting treatments. They can
do this without swapping any patient records. Keeping the data within their
own walls allows healthcare providers to tap into bigger datasets while making
sure sensitive medical info stays protected. This approach is super handy in
medical imaging. With FL, diagnostic models can be trained using data col-
lected from various healthcare spots. This boosts accuracy and outcomes while
safeguarding privacy.
In finance, FL is a really helpful tool for managing risk & catching fraud. Banks,
credit institutions, and financial service providers can work together to train ma-
chine learning models for things like credit scoring or figuring out loan risks. They
do this without needing to share any customer info. This way, they can combine
their knowledge & make their models better while still following strict financial
rules. By allowing secure teamwork, FL makes decision-making better in fi-
nance and helps drive innovation without putting individual client privacy at
risk.
The IoT sector is another place were FL shines brightly. IoT gadgets like sen-
sors, wearables, and smart home devices create tons of data. A lot of it is to spread
out or sensitive to just put in one place. FL lets these devices train models locally
and only share updates about those models. This means it’s perfect for making
devices work better while keeping data private. For instance, in smart homes, FL
can help save energy or enhance personalized services without sending user data
to a central hub. In industrial IoT, FL can be useful for predictive maintenance.
Here, sensors on machines team up to forecast equipment failures, boosting effi-
ciency & cutting down downtime. Federated Learning (FL) mixed with edge com-
puting makes it easier to work in situations where quick decisions are needed. In
edge computing, devices can do their work right there, without always needing to
connect to a main server. This is super important for things like self-driving cars,
which need to analyze data in real-time.
With FL, these cars can get better at making choices. They can train their mod-
els using local data like sensor info or traffic updates without putting sensitive data
at risk. Keeping that info safe is a big deal. FL is a practical answer for many fields
like healthcare, finance, the Internet of Things (IoT), and self-driving cars. Its way
of handling data helps keep it secure while still allowing for teamwork in develop-
ing models. Plus, it tackles big issues like privacy, how well systems talk to each
other, and the heavy lifting of computations. As research moves forward, we’ll
likely see FL grow in different areas. It’s becoming a must-have tech for businesses
that care about privacy and want to use their data smartly.

DOI: 10.4236/oalib.1112239 13 Open Access Library Journal

 N. E. M. R. Agripina et al.

7. Innovations and Advantages of Federated Learning

Federated Learning (FL) brings big changes and perks compared to the usual ma-
chine learning (ML) methods, especially when it comes to privacy, efficiency, and
scalability. With FL, you don’t need to store data in one spot with traditional ML.
Instead, it lets models train on data spread out across different places. Traditional
ML requires gathering all the data into one location for training, which can
heighten the risk of data breaches. But FL keeps raw data on its original devices or
servers. So, only model updates get sent out. This helps keep things more private
since personal data stays put.
Another cool thing about FL is how it keeps things private. It’s got neat features
like differential privacy and homomorphic encryption built in, which means no-
body can trace data back to its specific user. This allows sensitive information to
be used for training without putting anyone at risk. Because of these privacy fea-
tures, FL works great in fields were keeping info safe is super important, like
healthcare or finance and even IoT (Internet of Things). In these areas, FL helps
train models while keeping personal data locked up tight.
FL also cuts down on how much data needs to be sent around. Traditional ML
usually requires a ton of data to be moved to a central spot for training, which
racks up high costs—especially when dealing with lots of info. FL sidesteps that
by just sharing model parameters & updates instead of raw data. That really
slashes the amount of data flying across networks. So, it’s not just more efficient;
it’s also better when there are limitations on bandwidth or processing power—like
with smartphones or IoT devices.
Plus, FL shines when it comes to personalizing models. Typical ML models may
not adapt well to specific user behavior since they work with centralized datasets.
But with FL, you can build models tailored to individual users by training directly
on local user info. This is super handy for things like mobile keyboards or recom-
mendation systems where people’s preferences can be very different.
Collaboration between organizations gets a boost, too, with FL. Multiple groups
can work together to train a model without having to share their actual datasets.
Take healthcare: hospitals could train a shared model to predict patient outcomes
without sharing sensitive health records. This fosters teamwork between different
organizations and helps them make the most of available data for training while
staying within strict privacy rules like GDPR and HIPAA.
Also worth mentioning is how well FL handles different types of data. Tradi-
tional ML usually assumes that the info it’s working with is uniform & similar
(that means IID), but that doesn’t hold true in many real-life situations. On the
flip side, FL does just fine with non-IID data—where information varies from de-
vice to device—which makes its models better at handling diverse sources while
keeping privacy intact.
Finally, don’t overlook how well FL fits with edge computing! In edge compu-
ting setups, devices close to the network’s edge can perform calculations right
there, allowing models to update quickly without needing constant connections

DOI: 10.4236/oalib.1112239 14 Open Access Library Journal

 N. E. M. R. Agripina et al.

back to a main server. This speedy integration boosts training times and cuts down
on delays—a huge bonus for real-time applications, like self-driving cars or ana-
lyzing data instantly in IoT systems.

8. Security and Privacy Concerns

Federated Learning (FL) effectively resolves data privacy concerns in contempo-
rary machine learning by fundamentally altering data handling and processing
methods. In conventional machine learning (ML), it is necessary to centralize
data, which means that information from various sources is collected in one loca-
tion for model training. This aggregation heightens the risk of data breaches, un-
authorized access, and misuse, particularly in cases involving sensitive personal
or organizational data. On the contrary, FL facilitates model training directly on
dispersed data sources without requiring the centralization of raw data. This de-
centralized approach guarantees that sensitive information remains on local de-
vices, thus minimizing the potential for exposure.
One major benefit of FL is its preservation of local data ownership. With cen-
tralized systems, once data is sent to a server, individuals or organizations may
lose control over its usage. FL addresses this issue by ensuring that data stays on
local devices. Instead of sharing the data itself, model updates are sent to a central
server, which enables individuals or organizations to engage in collaborative
learning while maintaining authority over their private information. This aspect
makes FL well-suited to data protection regulations such as GDPR, which focus
on user consent and data security.
The privacy aspect in FL is further enhanced through techniques like differen-
tial privacy and homomorphic encryption. Traditional ML approaches may still
allow sensitive information to be inferred from anonymized datasets. FL reduces
this risk by introducing noise into model updates through differential privacy, en-
suring that no single data entry can be traced back to its source. Homomorphic
encryption enables computations to occur on encrypted data, meaning model up-
dates can be compiled without decryption, thus ensuring that even the central
server cannot access the original data from individual devices.
FL also provides protection against data poisoning and model poisoning at-
tacks, which can significantly undermine traditional centralized ML systems. In
centralized models, an attacker can inject harmful data into the system, leading to
model corruption. Conversely, with FL, data remains decentralized, and individ-
ual contributions are given less weight, thus minimizing the potential impact of
an attack. Furthermore, secure aggregation techniques and differential privacy
methods can identify and mitigate the consequences of poisoned model updates,
safeguarding the global model’s integrity.
Another notable advantage of FL is its capacity to comply with stringent data
privacy regulations. In traditional ML, consolidating data across regions or into a
single site can breach data residency requirements, complicating adherence to lo-
cal privacy laws. FL naturally avoids these challenges by retaining data locally,

DOI: 10.4236/oalib.1112239 15 Open Access Library Journal

 N. E. M. R. Agripina et al.

enabling organizations to train models across different areas without violating le-
gal obligations. This is particularly advantageous for sectors like healthcare and
finance, where data privacy is a regulatory imperative.
FL also reduces the risk linked to single points of failure, which represent a sig-
nificant vulnerability in centralized ML systems. In conventional setups, a breach
of the central server can expose all the stored data. In FL, data is distributed
over multiple nodes, so even if one participant is compromised, the breach is
confined and limited. This greatly lowers the overall risk of a large-scale data
breach.

8.1. Limitations and Challenges of Federated Learning

Federated Learning (FL) possesses the potential to revolutionize decentralized
model training while safeguarding data privacy; however, it also encounters sev-
eral limitations. These challenges can hinder its efficacy, particularly in cases
of data imbalance, the presence of malicious nodes, and communication barri-
ers. A comprehensive examination of these issues is essential to grasp the prac-
tical concerns that must be tackled to fully exploit FL’s capabilities in real-
world scenarios.

8.1.1. Data Imbalance and Heterogeneity

A key limitation of FL is its sensitivity to data imbalance and heterogeneity among
various distributed devices or nodes. In numerous FL applications, the data across
different devices is not independently distributed and is identically distributed
(non-IID). This differs from conventional machine learning methods, where
models are typically trained on datasets that are homogeneously distributed. Data
imbalance occurs when some nodes possess more data or their data more accu-
rately reflects specific classes or characteristics, leading to skewed model updates.
In such situations, FL models might find it challenging to generalize effectively
across all nodes, resulting in biased predictions. For instance, if an FL framework
allows devices to collectively contribute data for a shared predictive model, but
one device has an excessive amount of data related to healthy patients compared
to sick ones (as seen in healthcare applications), the resulting model may be biased
towards the majority class. Such imbalances can significantly undermine the per-
formance of the overall model, particularly in instances where data from specific
devices dominate or where certain devices do not contribute adequately due to
smaller datasets or underrepresented features.
To combat this issue, researchers have suggested approaches such as weighted
aggregation to mitigate the effects of imbalanced data on overall model perfor-
mance. Nonetheless, these strategies are still under development and may not
completely resolve the complications caused by highly skewed data distributions.

8.1.2. Malicious Node Participation

Another major drawback of FL is its vulnerability to harmful participation by
nodes. In the FL setup, individual devices (also called clients or nodes) locally train

DOI: 10.4236/oalib.1112239 16 Open Access Library Journal

 N. E. M. R. Agripina et al.

models and send updates to a central server for aggregation. This decentralized
framework makes it possible for malicious entities to disrupt the model training
process by submitting erroneous, poisoned, or manipulated updates.
Malicious nodes can carry out various types of attacks, including:
 Data Poisoning: A malicious node can insert biased or false data into the train-
ing pipeline, altering the local model updates sent to the server. This could lead
to the global model converging inaccurately, resulting in flawed or prejudiced
predictions. For example, in a financial modeling context, a malicious partici-
pant might introduce fraudulent data that skews the model towards favorable
predictions for certain behaviors.
 Model Poisoning: Rather than tampering with the data itself, a harmful node
can modify the local model parameters before submitting them to the server,
thereby injecting detrimental updates into the global model. This action can
compromise the integrity of the overall model and potentially lead to wide-
spread system failures.
 Backdoor Attacks: Malicious participants can also embed backdoors within
the global model, which may remain inactive until a specific trigger occurs.
This poses significant security risks since attackers can activate these back-
doors after the model is deployed, causing the model to operate unpredictably
or harmfully.
While secure aggregation techniques and differential privacy are commonly
employed to alleviate some of these risks, the challenges remain considerable.
Cryptographic strategies like secure multi-party computation can help ensure that
model updates are kept private and unaltered, but they can be resource-intensive
and may introduce latency. Furthermore, identifying and eliminating malicious
updates in a decentralized environment—without direct access to original data—
continues to be a challenge, as current methods might inadvertently dismiss legit-
imate updates or fail to detect sophisticated attacks.

8.1.3. Communication Overhead and Scalability

The frequent communication required between distributed nodes and a central
server in FL also imposes limitations, particularly in environments with limited
resources. During each training round, devices must share their model updates,
consuming substantial network bandwidth, especially when dealing with large
models or a high number of participants. This situation is especially troublesome
in cases involving IoT devices or mobile networks, where communication is re-
stricted by bandwidth or energy limitations.
Although strategies such as model update compression, sparsification, and
asynchronous communication can help minimize communication costs, these so-
lutions may compromise model precision or delay convergence. Additionally, as
the number of participating devices increases, communication expenses can be-
come untenable, making FL less viable in heavily distributed environments lack-
ing proper infrastructure.

DOI: 10.4236/oalib.1112239 17 Open Access Library Journal

 N. E. M. R. Agripina et al.

8.1.4. Computational Load on Edge Devices

FL places significant computational demands on edge devices (for example,
smartphones and IoT sensors), which often possess limited processing power,
memory, and battery life. Unlike traditional machine learning that relies on cen-
tralized servers with robust computing capabilities, FL necessitates local compu-
tations on devices with restricted resources, potentially slowing the training pro-
cess or limiting the complexity of the models that can be sustained.
In a healthcare scenario, where mobile devices are employed to train local mod-
els using patient data, the constrained computational abilities of these devices may
hinder their capacity to handle large datasets or advanced neural networks. This
limitation not only affects the sophistication of deployable models but also ex-
tends the time needed to achieve an acceptable level of accuracy. Enhancing FL to
perform effectively on resource-constrained devices is an ongoing research focus,
but existing solutions often require trade-offs between efficiency, model complex-
ity, and accuracy.

8.1.5. Privacy vs. Accuracy Trade-Offs

FL also presents a conflict between privacy and model accuracy. Techniques like
differential privacy, which add noise to data or model updates to protect individ-
ual data points from being traced back to users, can diminish the performance of
the overall model. While privacy-preserving strategies are essential for compli-
ance with data protection laws, they frequently come with the downside of de-
creased model accuracy or slower convergence, especially in cases where data is
already scarce or noisy.
Methods such as homomorphic encryption and secure aggregation aim to
maintain privacy without sacrificing model efficacy; however, they come with
substantial computational and communication overheads. Striking a balance be-
tween privacy needs and the demand for high-accuracy models remains a critical
challenge for FL, particularly in sensitive domains like healthcare or autonomous
driving, where precision is crucial.

9. Challenges and Future Research

9.1. Future Directions for Federated Learning (FL)
The future of Federated Learning (FL) is focused on overcoming its existing lim-
itations, enhancing performance, scalability, security, and broadening its applica-
tions in various fields. Efforts are being directed toward tackling issues such as
data variability, potential malicious node involvement, communication efficiency,
and privacy issues while also seeking innovative technical solutions to improve
the FL framework. Here are some of the major pathways and suggested technical
approaches for FL.

9.2. Managing Data Imbalance and Non-IID Data

A major challenge facing Federated Learning (FL) is the non-identically distributed

DOI: 10.4236/oalib.1112239 18 Open Access Library Journal

 N. E. M. R. Agripina et al.

(non-IID) nature of data across multiple clients. Different devices typically gather
data that varies in distribution, which can lead to biased models if not appropri-
ately managed. To address this, researchers are investigating personalized feder-
ated learning, which involves developing customized models tailored to specific
groups of clients with similar data distributions or behaviors, rather than creating
a single global model. This method enables FL to provide more relevant models
to individual clients while still leveraging shared learning across the network. An-
other proposed solution is federated meta-learning, or “learning to learn,” de-
signed to enhance the global model’s adaptability to diverse data distributions.
This could help the global model quickly adjust to new trends and variations in
client data. Additionally, the creation of dynamic aggregation strategies can im-
prove FL’s efficiency by weighting updates according to data quality or represent-
ativeness instead of treating all updates uniformly as in current methods like Fed-
erated Averaging.

9.3. Resilience Against Malicious Participants and Secure

Aggregation
Strengthening the robustness of FL against adversarial threats, such as data poi-
soning and model manipulation, is a vital area for future investigation. Malicious
participants might submit incorrect or altered updates, jeopardizing the integrity
of the global model. It is crucial to implement Byzantine-resilient algorithms that
can identify and counter harmful updates from compromised nodes. Algorithms
like Krum, Bulyan, and Trimmed Mean concentrate on aggregating the most
trustworthy updates while ignoring outliers that may have been tampered with,
though these techniques still require optimization for large-scale applications.
Verifiable federated learning presents a promising alternative since it allows cli-
ents to authenticate the legitimacy of their model updates without disclosing sen-
sitive information. The incorporation of zero-knowledge proofs (ZKPs) could en-
sure that servers can validate update correctness while safeguarding client privacy.
Moreover, establishing incentive mechanisms and reputation systems may foster
a trustworthy environment in FL by rewarding honest engagement and penalizing
dishonest behavior. The integration of blockchain technology could also provide
a secure ledger of updates, strengthening system dependability and diminishing
the effect of compromised participants.

9.4. Enhancing Communication and Scalability

In FL, substantial communication is necessary between clients and the central
server, especially in large-scale deployments, making it essential to minimize com-
munication overhead while preserving accuracy. One tactic for achieving this in-
volves model compression techniques, including model sparsification, quantiza-
tion, and gradient compression. These strategies can reduce the size of updates
sent between clients and servers, thereby accelerating training and decreasing
bandwidth usage. For instance, Top-K Gradient Compression only transmits the

DOI: 10.4236/oalib.1112239 19 Open Access Library Journal

 N. E. M. R. Agripina et al.

most significant gradients, which can lead to considerable cuts in communication

expenses. Another approach is the shift towards completely decentralized FL ar-
chitectures, often referred to as peer-to-peer FL. By eliminating the need for a
central server, devices can communicate directly with each other, sharing updates
and models in a distributed manner. This peer-to-peer communication model re-
moves central aggregation bottlenecks and enhances scalability. Additionally,
asynchronous federated learning permits devices to update the global model in-
dependently rather than at the same time, which is especially beneficial in scenar-
ios where numerous devices may be connected to the network intermittently,
thereby reducing the demand for continuous communication and enhancing FL’s
overall scalability.

9.5. Advancing Privacy-Preserving Techniques

While FL offers inherent privacy benefits by retaining data locally, certain vulner-
abilities still permit sensitive information to be inferred from model updates, par-
ticularly without the implementation of advanced privacy techniques like differ-
ential privacy or encryption. Homomorphic encryption serves as a potential solu-
tion, as it allows computations on encrypted data without needing decryption,
providing robust privacy assurances for FL. However, this method involves a sig-
nificant computational load, so future investigations should aim to improve the
efficiency of homomorphic encryption to render it more practical for real-time FL
applications. Differential privacy adds noise to model updates to prevent the leak-
age of sensitive information, although it can negatively affect model performance.
To counteract this, refined control measures in differential privacy should be cre-
ated to reduce the trade-off between privacy and accuracy by carefully modifying
noise levels based on data sensitivity. Secure aggregation protocols, such as multi-
party computation (MPC) and advanced aggregation techniques, should also be
enhanced to guarantee privacy-preserving updates in expansive networks while
maintaining system effectiveness.

9.6. Exploring New Sectors for FL Application

The potential applications of FL continue to expand as more sectors embrace de-
centralized machine learning; however, there remain numerous fields where FL
has not been thoroughly utilized. In areas such as healthcare, finance, and smart
cities, cross-silo FL could be particularly advantageous, especially where legal re-
strictions limit data sharing. This method enables different institutions with vari-
ous datasets but the same user base to collaborate on model training without
breaching privacy laws. Research into vertical federated learning, wherein datasets
are divided based on different feature sets, could further promote the adoption of
FL in these sectors. The convergence of FL with edge computing represents an-
other promising avenue, potentially facilitating the creation of real-time, low-la-
tency AI applications. Optimizing FL for edge settings necessitates improving
communication protocols for edge devices and ensuring system resilience against

DOI: 10.4236/oalib.1112239 20 Open Access Library Journal

 N. E. M. R. Agripina et al.

intermittent connectivity. Additionally, federated reinforcement learning—a rel-

atively innovative area that merges FL with reinforcement learning—holds con-
siderable promise in dynamic environments such as robotics, autonomous vehi-
cles, and smart grids, where agents learn by interacting with their environments,
paving the way for collaborative AI systems in real time.

9.7. Federated Learning as a Service (FLaaS)

The concept of providing FL as a service (FLaaS) is an emerging direction that
could boost the uptake of FL across various industries. By offering managed plat-
forms that enable organizations to implement FL models without the requirement
for in-house expertise, FLaaS could broaden access to this technology. For in-
stance, cloud service providers could incorporate FL frameworks into their exist-
ing services, managing secure aggregation, model oversight, and deployment,
thereby making it accessible to smaller organizations. Standardizing FL protocols
and frameworks is essential for this transition, as it would ensure interoperability
between various platforms and devices, simplifying the adoption process. Such
standardization will also make it easier for organizations to embrace FL technol-
ogy while mitigating the complexities involved in overseeing distributed learning
environments.

10. Conclusions
Federated learning is a novel approach in machine learning that allows decentral-
ized model training while preserving privacy and security of users’ data. By over-
coming inherent drawbacks of centralized learning, such as data aggregation, FL
provides a groundbreaking solution to privacy, regulatory compliance, and secu-
rity problems. Its benefit is the capacity to train models across devices or servers
without moving sensitive data, which appeals to areas with high privacy require-
ments, such as healthcare, finance, or the Internet of Things.
While federated learning has great potential, there are several challenges that
need to be overcome to unlock it fully. Firstly, one of the more critical challenges
is heterogeneous data, indicated by the varying differences in data distribution
across devices. This leads to problems with the generalization of global models
and most often results in the local model being biased. Second, there are signifi-
cant communication/computation costs that are also critical in the context of sys-
tems that depend on a high number of resource-constrained devices. Their limited
resources often create problems both in terms of communication and data pro-
cessing requirements. The necessity to frequently send updates to the central
server may lead to overextending the devices’ processing capabilities, while the
deficiencies in computational power may lead to certain limitations on the com-
plexity of the models that can be developed and deployed.
Besides, security concerns are of particular importance in FL. Since the FL set-
ting is decentralized, the system is subject to adversarial attacks, such as data or
model poisoning, whereby malicious participants corrupt the aggregation process

DOI: 10.4236/oalib.1112239 21 Open Access Library Journal

 N. E. M. R. Agripina et al.

to mislead the central server about the true global model. The resilience of FL
systems against such threats and application of secure aggregation techniques,
which guarantee that security and privacy of the training data are maintained, are
much required for the safe and reliable usage of the approach.
As a preliminary, it must be mentioned that to overcome the outlined chal-
lenges, further research should be concerted around several core aspects. In par-
ticular, the considerable extension of the existing federated optimization tech-
niques and the dynamic aggregation strategies would help to solve the problems
of data heterogeneity and enhance the model’s performance. Moreover, more ef-
fective privacy-preserving mechanisms, such as homomorphic encryption and
differential privacy, should be developed to ensure the optimal level of privacy and
high model accuracy. Lastly, more efficient communication protocols, such as
model compression, and more effective approaches to asynchronous updates,
could make the FL more scalable and less resource-consuming for larger deploy-
ments.
To conclude, federated learning presents a viable solution to collaborative ma-
chine learning in privacy-preserving and distributed settings. As the field ad-
vances, overcoming some of its existing constraints will allow it to become a sig-
nificant part of future AI-driven applications, enabling industries to harness their
data’s potential across the board without compromising security or privacy. Fur-
ther progression in the area will be allowed by academia, industry, and policymak-
ers alike, promoting the synthesis of decentralized AI systems in a variety of fields.

Conflicts of Interest
The author declares no conflicts of interest.

References
[1] McMahan, H.B. and Moore, E. (2016) Communication-Efficient Learning of Deep
Networks from Decentralized Data. Proceedings of the 20th International Confer-
ence on Artificial Intelligence and Statistics (AISTATS), Fort Lauderdale, 1273-1282.
[2] He, K., Zhang, Y., Ren, S. and Sun, J. (2020) Securing Federated Learning against
Malicious Clients. Proceedings of the IEEE Symposium on Security and Privacy
(S&P), 1-16.
[3] Yang, Q., Liu, Y., Tian, Z., Yu, S. and Chen, K. (2020) Federated Learning: Challenges,
Methods, and Future Directions. IEEE Transactions on Parallel and Distributed Sys-
tems, 30, 1799-1819.
[4] Ho, Q.B., Cichocki, A. and Hong, T.P. (2013) Learning Processes in Decentralized
Collaborative Working Environments. International Conference on Industrial Tech-
nology, 793-798.
[5] Beltrán, E.T.M., Pérez, M.Q., Sánchez, P.M.S., Bernal, S.L., Bovet, G., Pérez, M.G.,
Pérez, G.M. and Celdrán, A.H. (2023) Decentralized Federated Learning: Fundamen-
tals, State of the Art, Frameworks, Trends, and Challenges. IEEE Communications
Surveys & Tutorials, 25, 2983-3013. https://doi.org/10.1109/COMST.2023.3315746
[6] Shaheen, M., Farooq, M.S. and Umer, T. (2024) AI-Empowered Mobile Edge Compu-
ting: Inducing Balanced Federated Learning Strategy over Edge for Balanced Data

DOI: 10.4236/oalib.1112239 22 Open Access Library Journal

 N. E. M. R. Agripina et al.

and Optimized Computation Cost. Journal of Cloud Computing, 13, Article No. 52.
https://doi.org/10.1186/s13677-024-00614-y
[7] Farooq, U., Naseem, S., Li, J., Mahmood, T., Rehman, A., Saba, T. and Mustafa, L.
(2023) Analysis of the Factors Influencing the Predictive Learning Performance Us-
ing Federated Learning. Preprint. https://doi.org/10.21203/rs.3.rs-3243194/v1
[8] Mothukuri, V., Parizi, R.M., Pouriyeh, S., Huang, Y., Dehghantanha, A. and Sri-
vastava, G. (2021) A Survey on Security and Privacy of Federated Learning. Future
Generation Computer Systems, 115, 619-640.
https://doi.org/10.1016/j.future.2020.10.007
[9] Pasquini, C. and Böhme, R. (2020) Trembling Triggers: Exploring the Sensitivity of
Backdoors in DNN-Based Face Recognition. EURASIP Journal on Information Se-
curity, 2020, Article No. 12. https://doi.org/10.1186/s13635-020-00104-z
[10] Truex, S., Elsabrouty, M., Mhamdi, L., Felber, P. and Raynal, M. (2019) Hybrid-One:
Enhancing Privacy and Utility in Federated Learning with Hybrid-Privacy Strategies.
IEEE Transactions on Parallel and Distributed Systems, 31, 1736-1749.
[11] Smith, J. and Sekar, R. (2019) An Analysis of Fuzzy Logic Applications in Industrial
Engineering: Current Use and Future Prospects. Journal of Industrial Engineering
Research, 15, 45-58.
[12] Smith, G. and Sekar, V. (2019) Federated Learning: A Privacy-Preserving Collabora-
tive Machine Learning Framework. Proceedings of the ACM International Confer-
ence on Knowledge Discovery and Data Mining (KDD), 78-87.
[13] Nasr, M., Shokri, R. and Houmansadr, A. (2019) Comprehensive Privacy Analysis of
Deep Learning: Passive and Active White-Box Inference Attacks against Centralized
and Federated Learning. 2019 IEEE Symposium on Security and Privacy (SP), San
Francisco, 19-23 May 2019, 739-753. https://doi.org/10.1109/sp.2019.00065
[14] Cai, J. and Venkatasubramanian, K. (2018) Anomaly Detection for Wearable Health
Monitoring Systems: A K-Nearest Neighbor Model with Kernel Density Estimation.
Proceedings of the 2018 IEEE International Conference on Big Data (Big Data), 4488-
4493.

DOI: 10.4236/oalib.1112239 23 Open Access Library Journal