Federated Learning as a Privacy Solution - An Overview

Abstract

The Fourth Industrial Revolution suggests smart and automated industrial solutions by incorporating Artificial Intelligence into it. Today, the world of technology is highly dependent on Machine Learning (ML) and Deep Learning (DL) and their applications. All these ML/DL models, which bring huge benefits and provide Industry 4.0 solutions, require a bulk of data, extensive computational power, and storage for enhanced performance and accuracy. With the current jurisdictions on privacy all over the world, it is hard to access the required amount of data without giving the data ownership to the centralized silos. Taking model to the data source is the idea that makes Federated Learning (FL) a unique and better-suited solution in this situation. In this paper, we present a review of FL, its learning models, aggregation algorithms, frameworks, and the challenges faced by this new paradigm of decentralized and distributed Machine Learning. We discuss the potential applications of FL in various domains that can help improve the efficiency and flexibility of industrial processes. We also talk about their impact on changing the model training trends altogether in terms of data privacy, decentralization, security, and resource management. The main contribution of this work is to provide a comprehensive and concise review and comparative analysis of various frameworks and aggregation algorithms, followed by a discussion of challenges currently faced by FL.

Full text

ScienceDirect
Available online at www.sciencedirect.com
Procedia Computer Science 217 (2023) 316–325
1877-0509 © 2022 The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0)
Peer-review under responsibility of the scientic committee of the 4th International Conference on Industry 4.0 and Smart
Manufacturing
10.1016/j.procs.2022.12.227
10.1016/j.procs.2022.12.227 1877-0509
© 2022 The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0)
Peer-review under responsibility of the scientic committee of the 4th International Conference on Industry 4.0 and Smart
Manufacturing
Available online at www.sciencedirect.com
Procedia Computer Science 00 (2022) 000–000
www.elsevier.com/locate/procedia
4th International Conference on Industry 4.0 and Smart Manufacturing
Federated Learning as a Privacy Solution - An Overview
Mashal Khana,∗
, Frank G. Glavina, Matthias Nicklesa
aSchool of Computer Science, University of Galway, Ireland
Abstract
The Fourth Industrial Revolution suggests smart and automated industrial solutions by incorporating Artificial Intelligence into it.
Today, the world of technology is highly dependent on Machine Learning (ML) and Deep Learning (DL) and their applications. All 8
these ML/DL models, which bring huge benefits and provide Industry 4.0 solutions, require a bulk of data, extensive computational 9
power, and storage for enhanced performance and accuracy. With the current jurisdictions on privacy all over the world, it is hard 10
to access the required amount of data without giving the data ownership to the centralized silos. Taking model to the data source is 11
the idea that makes Federated Learning (FL) a unique and better-suited solution in this situation. In this paper, we present a review 12
of FL, its learning models, aggregation algorithms, frameworks, and the challenges faced by this new paradigm of decentralized 13
and distributed Machine Learning. We discuss the potential applications of FL in various domains that can help improve the 14
efficiency and flexibility of industrial processes. We also talk about their impact on changing the model training trends altogether 15
in terms of data privacy, decentralization, security, and resource management. The main contribution of this work is to provide 16
a comprehensive and concise review and comparative analysis of various frameworks and aggregation algorithms, followed by a 17
discussion of challenges currently faced by FL.
©2022 The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the scientific committee of the 4th International Conference on Industry 4.0 and Smart Manu-
facturing.
Keywords: Federated Learning; Privacy Preservation; Decentralized and Distributed Machine Learning; Security
1. Introduction
With rapid advancements in Artificial Intelligence technologies, Machine Learning (ML) and in particular Deep
Learning (DL) are increasingly found in common IT applications. These applications range from medical, engineering,
and the Internet of Things (IoT), to marketing and business analytic tools. Although this advancement in technology
is beneficial and a big step towards the 4th Industrial Revolution, these algorithms require vast amounts of data
for training and testing the models. Acquiring that data is easy in some cases, however, when the models require
training on user data, privacy issues become a major concern. Various regulations and legal policies are introduced for
the preservation of user privacy such as the General Data Protection Regulation (GDPR) [14] and Health Insurance
∗Corresponding author. Tel.: +353 89 209 5203 ;
E-mail address: m.khan18@nuigalway.ie
1877-0509 ©2022 The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the scientific committee of the 4th International Conference on Industry 4.0 and Smart Manufacturing.
Available online at www.sciencedirect.com
Procedia Computer Science 00 (2022) 000–000
www.elsevier.com/locate/procedia
4th International Conference on Industry 4.0 and Smart Manufacturing
Federated Learning as a Privacy Solution - An Overview
Mashal Khana,∗
, Frank G. Glavina, Matthias Nicklesa
aSchool of Computer Science, University of Galway, Ireland
Abstract
The Fourth Industrial Revolution suggests smart and automated industrial solutions by incorporating Artificial Intelligence into it.
Today, the world of technology is highly dependent on Machine Learning (ML) and Deep Learning (DL) and their applications. All 8
these ML/DL models, which bring huge benefits and provide Industry 4.0 solutions, require a bulk of data, extensive computational 9
power, and storage for enhanced performance and accuracy. With the current jurisdictions on privacy all over the world, it is hard 10
to access the required amount of data without giving the data ownership to the centralized silos. Taking model to the data source is 11
the idea that makes Federated Learning (FL) a unique and better-suited solution in this situation. In this paper, we present a review 12
of FL, its learning models, aggregation algorithms, frameworks, and the challenges faced by this new paradigm of decentralized 13
and distributed Machine Learning. We discuss the potential applications of FL in various domains that can help improve the 14
efficiency and flexibility of industrial processes. We also talk about their impact on changing the model training trends altogether 15
in terms of data privacy, decentralization, security, and resource management. The main contribution of this work is to provide 16
a comprehensive and concise review and comparative analysis of various frameworks and aggregation algorithms, followed by a 17
discussion of challenges currently faced by FL.
©2022 The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the scientific committee of the 4th International Conference on Industry 4.0 and Smart Manu-
facturing.
Keywords: Federated Learning; Privacy Preservation; Decentralized and Distributed Machine Learning; Security
1. Introduction
With rapid advancements in Artificial Intelligence technologies, Machine Learning (ML) and in particular Deep
Learning (DL) are increasingly found in common IT applications. These applications range from medical, engineering,
and the Internet of Things (IoT), to marketing and business analytic tools. Although this advancement in technology
is beneficial and a big step towards the 4th Industrial Revolution, these algorithms require vast amounts of data
for training and testing the models. Acquiring that data is easy in some cases, however, when the models require
training on user data, privacy issues become a major concern. Various regulations and legal policies are introduced for
the preservation of user privacy such as the General Data Protection Regulation (GDPR) [14] and Health Insurance
∗Corresponding author. Tel.: +353 89 209 5203 ;
E-mail address: m.khan18@nuigalway.ie
1877-0509 ©2022 The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the scientific committee of the 4th International Conference on Industry 4.0 and Smart Manufacturing.
2Author name /Procedia Computer Science 00 (2022) 000–000
Portability and Accountability Act (HIPAA) [20]. Also, keeping in mind the increasing use of AI, the European
Commission is planning on introducing a legal framework for AI and the associated risks [41].
Some issues with the current ML model training:
Limited data availability: Gathering data to a central source can be expensive.
User data privacy and security: Single sources of sensitive data could be compromised.
Centralized and expensive resources for training: Extensive compute and power resources required.
With the increased user awareness of privacy concerns and the associated laws, users are concerned about the security
of their data and their privacy and are often reluctant to share their data. The accuracy of the Machine Learning
models is then negatively affected due to the lack of data availability. The centralized training of ML/DL models
requires huge amounts of data as well as powerful computational resources. Arranging enough energy and power for
the training and testing of these models is another hindrance to the deployment of these algorithms. Also, the data
collected from various sources are in different formats, which requires a lot of pre-processing and cleaning before it
is used to train a model. Due to the above-mentioned reasons, ML is restricted from reaching its full potential for
industrial applications in some cases [38,31]. Federated Learning (FL), or Federated Machine Learning (FedML),
was introduced as a solution to all the above-mentioned problems.
The standard systematic literature review methodology is followed for this study. Different search filters and queries
were used to filter out the relevant literature from the related databases i.e. Google Scholar, Semantic Scholar, Science
Direct, IEEE Xplore Digital Library, Research Gate etc. To make sure only the most related literature is filtered,
specific keywords and logical operators were used e.g. ”Federated Machine Learning”+”client selection”. Date filters
were also used to access the most recent work. Most of the work presented in this study is done after 2016, around 70%
of the literature is published 2019 afterwards. Almost 50% of the work presented in this survey is by different research
groups from big tech companies like Google, Microsoft and IBM, while the rest is by independent researchers and
research centers around the world.
In the next section, we introduce FL and its basic workings. In Section 3, we present a taxonomy of FL categorization
based on different factors. Section 4 is about ensuring privacy in FL and the major techniques used. Section 5 outlines
the key considerations for FL. Section 6 summarises the application domains of FL and highlights some successful
projects. In Section 7, we discuss the issues and challenges currently faced by FL before providing some concluding
remarks in Section 8.
2. Federated Learning
Revolutionizing the industry by shifting the computation to the edge devices rather than servers, saving the planet
from over-heating due to massive energy consumption at data centers, ensuring privacy and data security, and building
trust among the communicating parties are the main goals of FL.
The concept works as follows: Instead of collecting data into a single source to train the model, the model is sent
to the respective data sources for training. The data source returns the weighted graphs, or results of the training, to
the central server. This helps to reduce privacy and security issues and the model can be trained efficiently. It has been
shown that better results can be achieved with having access to a diverse and larger set of data from various resources,
as the data privacy concerns are addressed by FL [62,6,28,44].
The idea was first presented by Google in 2016 and was later implemented on GBoard for Android in 2017 [63].
The authors trained an FL model on the data of individual edge (mobile) devices. The results were shared in the
form of vectors/graphs with the central cloud server to ensure data privacy. The updates from all users were averaged
securely using Secure Aggregation Protocol [7], making sure that only the averaged updates can be accessed and not
the individual updates from a single user. To make the system efficient, and to avoid processing speed and memory
issues, it was made sure that the training is done only when the device is in an idle state and on a free WiFi connection.
The updates sent were also compressed to avoid any kind of disruptions in the communication [17].
Any FL model follows the below-mentioned steps:
1. The server shares the model with a set of clients/peers.
2. The clients train the model locally with their device/centre’s data.
3. Each client sends the encrypted gradients to the server.
4. The server aggregates the gradients from all clients.
5. The global model is updated and shared with the clients in the next iteration.

Mashal Khan et al. / Procedia Computer Science 217 (2023) 316–325 317
Available online at www.sciencedirect.com
Procedia Computer Science 00 (2022) 000–000
www.elsevier.com/locate/procedia
4th International Conference on Industry 4.0 and Smart Manufacturing
Federated Learning as a Privacy Solution - An Overview
Mashal Khana,∗
, Frank G. Glavina, Matthias Nicklesa
aSchool of Computer Science, University of Galway, Ireland
Abstract
The Fourth Industrial Revolution suggests smart and automated industrial solutions by incorporating Artificial Intelligence into it.
Today, the world of technology is highly dependent on Machine Learning (ML) and Deep Learning (DL) and their applications. All 8
these ML/DL models, which bring huge benefits and provide Industry 4.0 solutions, require a bulk of data, extensive computational 9
power, and storage for enhanced performance and accuracy. With the current jurisdictions on privacy all over the world, it is hard 10
to access the required amount of data without giving the data ownership to the centralized silos. Taking model to the data source is 11
the idea that makes Federated Learning (FL) a unique and better-suited solution in this situation. In this paper, we present a review 12
of FL, its learning models, aggregation algorithms, frameworks, and the challenges faced by this new paradigm of decentralized 13
and distributed Machine Learning. We discuss the potential applications of FL in various domains that can help improve the 14
efficiency and flexibility of industrial processes. We also talk about their impact on changing the model training trends altogether 15
in terms of data privacy, decentralization, security, and resource management. The main contribution of this work is to provide 16
a comprehensive and concise review and comparative analysis of various frameworks and aggregation algorithms, followed by a 17
discussion of challenges currently faced by FL.
©2022 The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the scientific committee of the 4th International Conference on Industry 4.0 and Smart Manu-
facturing.
Keywords: Federated Learning; Privacy Preservation; Decentralized and Distributed Machine Learning; Security
1. Introduction
With rapid advancements in Artificial Intelligence technologies, Machine Learning (ML) and in particular Deep
Learning (DL) are increasingly found in common IT applications. These applications range from medical, engineering,
and the Internet of Things (IoT), to marketing and business analytic tools. Although this advancement in technology
is beneficial and a big step towards the 4th Industrial Revolution, these algorithms require vast amounts of data
for training and testing the models. Acquiring that data is easy in some cases, however, when the models require
training on user data, privacy issues become a major concern. Various regulations and legal policies are introduced for
the preservation of user privacy such as the General Data Protection Regulation (GDPR) [14] and Health Insurance
∗Corresponding author. Tel.: +353 89 209 5203 ;
E-mail address: m.khan18@nuigalway.ie
1877-0509 ©2022 The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the scientific committee of the 4th International Conference on Industry 4.0 and Smart Manufacturing.
2Author name /Procedia Computer Science 00 (2022) 000–000
Portability and Accountability Act (HIPAA) [20]. Also, keeping in mind the increasing use of AI, the European
Commission is planning on introducing a legal framework for AI and the associated risks [41].
Some issues with the current ML model training:
Limited data availability: Gathering data to a central source can be expensive.
User data privacy and security: Single sources of sensitive data could be compromised.
Centralized and expensive resources for training: Extensive compute and power resources required.
With the increased user awareness of privacy concerns and the associated laws, users are concerned about the security
of their data and their privacy and are often reluctant to share their data. The accuracy of the Machine Learning
models is then negatively affected due to the lack of data availability. The centralized training of ML/DL models
requires huge amounts of data as well as powerful computational resources. Arranging enough energy and power for
the training and testing of these models is another hindrance to the deployment of these algorithms. Also, the data
collected from various sources are in different formats, which requires a lot of pre-processing and cleaning before it
is used to train a model. Due to the above-mentioned reasons, ML is restricted from reaching its full potential for
industrial applications in some cases [38,31]. Federated Learning (FL), or Federated Machine Learning (FedML),
was introduced as a solution to all the above-mentioned problems.
The standard systematic literature review methodology is followed for this study. Different search filters and queries
were used to filter out the relevant literature from the related databases i.e. Google Scholar, Semantic Scholar, Science
Direct, IEEE Xplore Digital Library, Research Gate etc. To make sure only the most related literature is filtered,
specific keywords and logical operators were used e.g. ”Federated Machine Learning”+”client selection”. Date filters
were also used to access the most recent work. Most of the work presented in this study is done after 2016, around 70%
of the literature is published 2019 afterwards. Almost 50% of the work presented in this survey is by different research
groups from big tech companies like Google, Microsoft and IBM, while the rest is by independent researchers and
research centers around the world.
In the next section, we introduce FL and its basic workings. In Section 3, we present a taxonomy of FL categorization
based on different factors. Section 4 is about ensuring privacy in FL and the major techniques used. Section 5 outlines
the key considerations for FL. Section 6 summarises the application domains of FL and highlights some successful
projects. In Section 7, we discuss the issues and challenges currently faced by FL before providing some concluding
remarks in Section 8.
2. Federated Learning
Revolutionizing the industry by shifting the computation to the edge devices rather than servers, saving the planet
from over-heating due to massive energy consumption at data centers, ensuring privacy and data security, and building
trust among the communicating parties are the main goals of FL.
The concept works as follows: Instead of collecting data into a single source to train the model, the model is sent
to the respective data sources for training. The data source returns the weighted graphs, or results of the training, to
the central server. This helps to reduce privacy and security issues and the model can be trained efficiently. It has been
shown that better results can be achieved with having access to a diverse and larger set of data from various resources,
as the data privacy concerns are addressed by FL [62,6,28,44].
The idea was first presented by Google in 2016 and was later implemented on GBoard for Android in 2017 [63].
The authors trained an FL model on the data of individual edge (mobile) devices. The results were shared in the
form of vectors/graphs with the central cloud server to ensure data privacy. The updates from all users were averaged
securely using Secure Aggregation Protocol [7], making sure that only the averaged updates can be accessed and not
the individual updates from a single user. To make the system efficient, and to avoid processing speed and memory
issues, it was made sure that the training is done only when the device is in an idle state and on a free WiFi connection.
The updates sent were also compressed to avoid any kind of disruptions in the communication [17].
Any FL model follows the below-mentioned steps:
1. The server shares the model with a set of clients/peers.
2. The clients train the model locally with their device/centre’s data.
3. Each client sends the encrypted gradients to the server.
4. The server aggregates the gradients from all clients.
5. The global model is updated and shared with the clients in the next iteration.

318 Mashal Khan et al. / Procedia Computer Science 217 (2023) 316–325
Author name /Procedia Computer Science 00 (2022) 000–000 3
Fig. 1: Steps of FL
3. Categorisation of Federated Learning
FL is divided into different categories based on various factors. The major categorization is based on data distribu-
tion across the edge devices/clients, the architecture, and model training.
Fig. 2: Categories of Federated Learning
3.1. Data Distribution
FL, while training models on data sources, consider the way data is distributed across them. There are two main
distributions which are cross device and cross silos. In cross device training of the model, the model is distributed
among the edge devices and is trained on that local data on each device. However, FL also works well for cross silos
distribution of the data i.e., the local model is trained at data centers/data silos and is aggregated into a global model
at a centralized point.
3.2. Learning Models
Another metric for the categorization of FL types is the learning model. The model can be trained in different
fashions such as Horizontal Federated Learning,Vertical Federated Learning, and Federated Transfer Learning.
Horizontal Federated Learning: The same feature space across all data sources is used. Google’s GBoard project
is a Cross-Device Horizontal FL working in a centralized fashion [17,19].
Vertical Federated Learning: This model is also called Feature Space Federated Learning. Different feature sets
across the data sources are used, the sample space may or may not be the same. Entity alignment and data modeling
techniques are required with Vertical FL, as feature sets are different across different data sources. FedAI Technology
4Author name /Procedia Computer Science 00 (2022) 000–000
Enabler FATE project is an example of Vertical FL. [16].
Federated Transfer Learning: It is like Vertical FL utilized with a pre-trained model that is trained on a similar
dataset for solving a different problem e.g., making a movie recommendation for the user based on their past browsing
behavior [8,25]. The FedHealth Project discussed in Sect. 6 uses Federated Transfer Learning.
3.3. Architecture
There are different modes in which an FL model can be trained based on architectural design and interactions
between different elements of the system. There are centralized and decentralized approaches to training a FL Model.
In the centralized approach there is a central server coordinating the communication with various data sources. The
central server is responsible for the selection of clients taking part in model training, aggregation of local models into
a global model, and communicating it among all parties. However, there is no central server for management and
coordination in the decentralized setting. Either the peers help and coordinate with each other or a blockchain [27]
can be used to facilitate the model sharing and aggregation of local models in a distributed setting.
4. Federated Learning as a Privacy Solution
The idea of FedML was suggested to incorporate privacy into ML models, shifting the computations to the edge
devices, and building trust among the communicating parties. Due to the increased awareness regarding data privacy
and the associated laws, it is necessary to maintain data security while training models. Privacy is added at multiple
levels and in different ways.
4.1. Cryptographic Solutions
Data privacy is assured due to the decentralization and distribution of data across multiple entities. FL does not
ask for user data, rather it sends the model to the data. But according to [70], user data can still be retrieved from
the gradients, thus compromising privacy. So, an additional layer of security can be added in the form of Differential
Privacy and Cryptographic algorithms such as Homomorphic Encryption and Secure Multiparty Computations (MPC)
[7,36,39]. Differential privacy adds noise to the data before it is processed thus hiding the actual data. But this comes
with a trade-offbetween privacy and the accuracy of the model. The addition of noise to the data can affect the
accuracy of the model. Cryptographic solutions encrypt the user data before it is used to train the local model. Even
after reversing the gradients, only encrypted data can be accessed which has no meaning. The decryption of results
only occurs at the server after the results from all the clients are aggregated, providing privacy and security. There is
no trade-offwith the model accuracy, but additional computations are required for the encryption and decryption of
the data and results.
4.2. Client Security
Clients selected for the training of the FL models play a crucial role in the security of data. In FL there is no need
for centralized storage of data. The data remains with the clients, and they can discard it after training the local model.
There are various client selection techniques suggested in the literature and some of them are listed in Section 5.3.
These techniques make sure that the clients selected are not malicious and are well-suited to train the model. The
sets of clients are selected according to the criteria specified in the selection technique. The set keeps on changing for
every iteration, deterring malicious entities to poison the data or model.
5. Key Considerations for Federated Learning
While designing an FL model, the following things should be kept in mind, for getting the maximum benefit of the
model and training it efficiently:
•Client selection techniques to ensure maximum responses from the end devices.
•Choosing the most suited aggregation algorithms.
•Framework selection according to the task at hand.

Mashal Khan et al. / Procedia Computer Science 217 (2023) 316–325 319
Author name /Procedia Computer Science 00 (2022) 000–000 3
Fig. 1: Steps of FL
3. Categorisation of Federated Learning
FL is divided into different categories based on various factors. The major categorization is based on data distribu-
tion across the edge devices/clients, the architecture, and model training.
Fig. 2: Categories of Federated Learning
3.1. Data Distribution
FL, while training models on data sources, consider the way data is distributed across them. There are two main
distributions which are cross device and cross silos. In cross device training of the model, the model is distributed
among the edge devices and is trained on that local data on each device. However, FL also works well for cross silos
distribution of the data i.e., the local model is trained at data centers/data silos and is aggregated into a global model
at a centralized point.
3.2. Learning Models
Another metric for the categorization of FL types is the learning model. The model can be trained in different
fashions such as Horizontal Federated Learning,Vertical Federated Learning, and Federated Transfer Learning.
Horizontal Federated Learning: The same feature space across all data sources is used. Google’s GBoard project
is a Cross-Device Horizontal FL working in a centralized fashion [17,19].
Vertical Federated Learning: This model is also called Feature Space Federated Learning. Different feature sets
across the data sources are used, the sample space may or may not be the same. Entity alignment and data modeling
techniques are required with Vertical FL, as feature sets are different across different data sources. FedAI Technology
4Author name /Procedia Computer Science 00 (2022) 000–000
Enabler FATE project is an example of Vertical FL. [16].
Federated Transfer Learning: It is like Vertical FL utilized with a pre-trained model that is trained on a similar
dataset for solving a different problem e.g., making a movie recommendation for the user based on their past browsing
behavior [8,25]. The FedHealth Project discussed in Sect. 6 uses Federated Transfer Learning.
3.3. Architecture
There are different modes in which an FL model can be trained based on architectural design and interactions
between different elements of the system. There are centralized and decentralized approaches to training a FL Model.
In the centralized approach there is a central server coordinating the communication with various data sources. The
central server is responsible for the selection of clients taking part in model training, aggregation of local models into
a global model, and communicating it among all parties. However, there is no central server for management and
coordination in the decentralized setting. Either the peers help and coordinate with each other or a blockchain [27]
can be used to facilitate the model sharing and aggregation of local models in a distributed setting.
4. Federated Learning as a Privacy Solution
The idea of FedML was suggested to incorporate privacy into ML models, shifting the computations to the edge
devices, and building trust among the communicating parties. Due to the increased awareness regarding data privacy
and the associated laws, it is necessary to maintain data security while training models. Privacy is added at multiple
levels and in different ways.
4.1. Cryptographic Solutions
Data privacy is assured due to the decentralization and distribution of data across multiple entities. FL does not
ask for user data, rather it sends the model to the data. But according to [70], user data can still be retrieved from
the gradients, thus compromising privacy. So, an additional layer of security can be added in the form of Differential
Privacy and Cryptographic algorithms such as Homomorphic Encryption and Secure Multiparty Computations (MPC)
[7,36,39]. Differential privacy adds noise to the data before it is processed thus hiding the actual data. But this comes
with a trade-offbetween privacy and the accuracy of the model. The addition of noise to the data can affect the
accuracy of the model. Cryptographic solutions encrypt the user data before it is used to train the local model. Even
after reversing the gradients, only encrypted data can be accessed which has no meaning. The decryption of results
only occurs at the server after the results from all the clients are aggregated, providing privacy and security. There is
no trade-offwith the model accuracy, but additional computations are required for the encryption and decryption of
the data and results.
4.2. Client Security
Clients selected for the training of the FL models play a crucial role in the security of data. In FL there is no need
for centralized storage of data. The data remains with the clients, and they can discard it after training the local model.
There are various client selection techniques suggested in the literature and some of them are listed in Section 5.3.
These techniques make sure that the clients selected are not malicious and are well-suited to train the model. The
sets of clients are selected according to the criteria specified in the selection technique. The set keeps on changing for
every iteration, deterring malicious entities to poison the data or model.
5. Key Considerations for Federated Learning
While designing an FL model, the following things should be kept in mind, for getting the maximum benefit of the
model and training it efficiently:
•Client selection techniques to ensure maximum responses from the end devices.
•Choosing the most suited aggregation algorithms.
•Framework selection according to the task at hand.

320 Mashal Khan et al. / Procedia Computer Science 217 (2023) 316–325
Author name /Procedia Computer Science 00 (2022) 000–000 5
5.1. Client Selection Techniques
Client Selection is a crucial aspect of model training. Some techniques for the selection of clients that are proposed
in literature are: Random Selection of Client that selects a set of clients randomly from a larger pool [46]. FedCS
also known as greedy selection, select the clients on the basis of their resources (computational and wireless channel
conditions) [40]. Incentives mechanism using Contract Theory urges the mobile devices with high-quality data to take
part in model training and overcome the issue of information asymmetry. It uses contract theory to reward the clients
based on their contribution [24]. Multi-Armed Bandit (MAB)-based client selection method is best client selection
technique when there is no information available about the computational and communication resources of the clients.
It helps in reducing the time consumption (by selection process) by balancing between the unknown clients and clients
with larger resources [64]. In Pulling Reduction with Local Compensation - PRLC clients pull the update-able global
model only when there is new or updated data at the data source [57]. All of them are used in practice and have
their pros and cons. Some of them are faster but do not have any specific criteria for shortlisting a set of clients,
e.g., Random Selection. Others consider different aspects like availability, resources of clients, and communication
frequency in the selection process to make the system better.
5.2. Aggregation Algorithms
The averaging/aggregation algorithms of FL ensure privacy and security by combining all the results into a global
model and keeping the local model updates and user data private. Table 1 provides an overview of important averaging
techniques in FL.
Table 1: Averaging Algorithm for Federated Learning
S.No Averaging
Algorithm Description
1.
Federated
Averaging –
FedAvg [8]
A few communication rounds are enough to train high quality models using FedAvg. The results
are demonstrated on various model architectures like a multi-layer perceptron, two different con-
volutional NNs, a two-layer character LSTM, and a large-scale word-level LSTM. Concepts like
Differential Privacy and Secure Multiparty Computations can be easily applied to FedAvg.
2.
Federated
Stochastic
Gradient Descent
– FedSGD [51]
This averaging technique is based on selective stochastic gradient descent and works on Neural
Networks. It provides privacy and control over learning objectives by using differential privacy.
3.
Federated
Personalization –
FedPer [2]
This technique addresses personalization by using a base +personalization layer approach for deep
feedforward neural networks. The effects of statistical heterogeneity can be dealt with using this
algorithm.
4.
Federated Match
Averaging –
FedMA [58]
This averaging algorithm presents a layer-wise matching and averaging of hidden elements with
similar feature extraction signatures for the construction of the global model. FedMA works well
for CNNs and LSTMs and resolves issues of data biasness.
5.
Federated
Learning with
Dynamic
Regularization –
FedDyn [1]
This averaging algorithm for neural networks chooses a subset of devices randomly, that are man-
aged by the server, in each round. A dynamic regularizer is proposed, to keep the device level and
global solutions aligned, at each round. This technique is different in terms of regularizing loss and
in attempting to parallelize gradient computations.
6.
Federated
Distribution –
FedDist [13]
Federated Distribution (FedDist) is a relatively new technique suggested for pervasive computing.
It modifies the model architecture by identifying dissimilarities between specific neurons (in the
case of DNNs) amongst the clients. Euclidean distance dissimilarity measurement is used, and
the client’s model is recognized because of non-IID data. It exhibits good experimental results
on different measures of generalization and personalization.
If the results from certain clients are not received in time, they are kept aside for the next iteration to avoid any
kind of delay, thus keeping the update process asynchronous and fast. The distributed nature of model training and
aggregation algorithms adds to the privacy of user data. The updates are incorporated asynchronously making the
6Author name /Procedia Computer Science 00 (2022) 000–000
training process efficient and fast as compared to the synchronous updates [7]. The first averaging technique was
FedSGD [51] by Google. They presented an upgraded version in the form of FedAvg [8] later that was based on model
averaging rather than gradient averaging and had a lower communication cost. Both techniques were shown to result
in poor performance at times. Some better techniques like FedMA [58] and FedPer [2] were presented, by IBM and
Adobe respectively. Some other techniques like Bayesian Nonparametric Federated Learning of Neural Networks
[65], Similarity based Federated Learning – SimFL [29], and Overlap-FedAvg [68] are also proposed in the literature.
5.3. Federated Learning Frameworks
There are various frameworks available for FL. Most of them are open source and provide almost all the func-
tionality required by FL, as previously discussed. Security is achieved using various techniques and the frameworks
support many ML models. Some of the main frameworks are discussed in the Table 2. Other frameworks are FATE
[59] , NVIDIA CLARA [11], Substra [15] and OpenFed [9] that provide support for various different use-cases.
Table 2: Federated Learning Frameworks
S.No Name Developer Description
1.
IBM Feder-
ated Learning
[35,45]
IBM Watson
Project
IBM FL was introduced as a part of the IBM Watson Project that is a part of IBM
Cloud Pak for Data and IBM Cloud Pak for Data as a Service. The framework
provides infrastructure and coordination for federated learning jobs to build
upon existing Deep Learning and Machine Learning algorithms. The frame-
work provides components for both the aggregator and the party (peers/clients).
Training of Neural Networks and Decision Trees is possible using this FL
framework.
2. TensorFlow
Federated [18]Google
It enables the developers to try the existing FL algorithms and experiment with
novel algorithms as well. It has two parts: 1) Federated Learning API, a high
level interface for the developers to use the existing FL for training and evalu-
ation and 2) Federated Learning Core, a lower level interface that can be used
for developing novel FL algorithms.
3.
PySyft +
PyGrid by Py-
Torch [43,42]
OpenMind
PySyft is used together with PyGrid to provide support for FL models. PyGrid
is a peer-to-peer platform where data owners can manage their data as well
as provide and monitor access to data clusters. PySyft allows users to perform
secure deep learning and is built as an extension of PyTorch, TensorFlow and
Keras.
4. Flower [5,4]
Joint project by
Oxford, UCL
and Cambridge
A comprehensive framework that supports large scale FL experiments on Het-
erogenous devices. The main features of this framework are scalability, Support
for heterogenous devices, Realism in results and Privacy preservation.
5. OpenFL [48] Intel Corp
Data Private Collaborative Framework for FL that can be extended to various
ML and DL models and support multiple aggregation algorithms like FedAvg,
FedProx, FedOpt etc.
6. Fed-BioMed
[21]
French Com-
puter Institute-
INRIA
This framework is mostly focused on Biomedical research and provides an FL
analysis framework. It is user friendly and easy to deploy.
7. XayNet [12] Xayn- Berlin It is a masked cross device FL Framework that supports Horizontal and Transfer
FL in both cross-device and cross-silos setting.
6. Applications of Federated Learning
The employment of smart devices and increasing automation in industrial solutions are the main goals of industry
4.0. Due to the decentralized and privacy-preserving nature of FL, it can be used in various fields to provide smart
solutions. An overview of some of the applications:
Healthcare: To keep up with the Industry 4.0 trends, there are research articles that encourage the use of FL in
healthcare and medical informatics due to its privacy-preserving nature. Data-driven medicines and treatments require
patients’ data that is confidential in nature. Thus, FL is the best solution for such cases [49]. FL also addresses many

Mashal Khan et al. / Procedia Computer Science 217 (2023) 316–325 321
Author name /Procedia Computer Science 00 (2022) 000–000 5
5.1. Client Selection Techniques
Client Selection is a crucial aspect of model training. Some techniques for the selection of clients that are proposed
in literature are: Random Selection of Client that selects a set of clients randomly from a larger pool [46]. FedCS
also known as greedy selection, select the clients on the basis of their resources (computational and wireless channel
conditions) [40]. Incentives mechanism using Contract Theory urges the mobile devices with high-quality data to take
part in model training and overcome the issue of information asymmetry. It uses contract theory to reward the clients
based on their contribution [24]. Multi-Armed Bandit (MAB)-based client selection method is best client selection
technique when there is no information available about the computational and communication resources of the clients.
It helps in reducing the time consumption (by selection process) by balancing between the unknown clients and clients
with larger resources [64]. In Pulling Reduction with Local Compensation - PRLC clients pull the update-able global
model only when there is new or updated data at the data source [57]. All of them are used in practice and have
their pros and cons. Some of them are faster but do not have any specific criteria for shortlisting a set of clients,
e.g., Random Selection. Others consider different aspects like availability, resources of clients, and communication
frequency in the selection process to make the system better.
5.2. Aggregation Algorithms
The averaging/aggregation algorithms of FL ensure privacy and security by combining all the results into a global
model and keeping the local model updates and user data private. Table 1 provides an overview of important averaging
techniques in FL.
Table 1: Averaging Algorithm for Federated Learning
S.No Averaging
Algorithm Description
1.
Federated
Averaging –
FedAvg [8]
A few communication rounds are enough to train high quality models using FedAvg. The results
are demonstrated on various model architectures like a multi-layer perceptron, two different con-
volutional NNs, a two-layer character LSTM, and a large-scale word-level LSTM. Concepts like
Differential Privacy and Secure Multiparty Computations can be easily applied to FedAvg.
2.
Federated
Stochastic
Gradient Descent
– FedSGD [51]
This averaging technique is based on selective stochastic gradient descent and works on Neural
Networks. It provides privacy and control over learning objectives by using differential privacy.
3.
Federated
Personalization –
FedPer [2]
This technique addresses personalization by using a base +personalization layer approach for deep
feedforward neural networks. The effects of statistical heterogeneity can be dealt with using this
algorithm.
4.
Federated Match
Averaging –
FedMA [58]
This averaging algorithm presents a layer-wise matching and averaging of hidden elements with
similar feature extraction signatures for the construction of the global model. FedMA works well
for CNNs and LSTMs and resolves issues of data biasness.
5.
Federated
Learning with
Dynamic
Regularization –
FedDyn [1]
This averaging algorithm for neural networks chooses a subset of devices randomly, that are man-
aged by the server, in each round. A dynamic regularizer is proposed, to keep the device level and
global solutions aligned, at each round. This technique is different in terms of regularizing loss and
in attempting to parallelize gradient computations.
6.
Federated
Distribution –
FedDist [13]
Federated Distribution (FedDist) is a relatively new technique suggested for pervasive computing.
It modifies the model architecture by identifying dissimilarities between specific neurons (in the
case of DNNs) amongst the clients. Euclidean distance dissimilarity measurement is used, and
the client’s model is recognized because of non-IID data. It exhibits good experimental results
on different measures of generalization and personalization.
If the results from certain clients are not received in time, they are kept aside for the next iteration to avoid any
kind of delay, thus keeping the update process asynchronous and fast. The distributed nature of model training and
aggregation algorithms adds to the privacy of user data. The updates are incorporated asynchronously making the
6Author name /Procedia Computer Science 00 (2022) 000–000
training process efficient and fast as compared to the synchronous updates [7]. The first averaging technique was
FedSGD [51] by Google. They presented an upgraded version in the form of FedAvg [8] later that was based on model
averaging rather than gradient averaging and had a lower communication cost. Both techniques were shown to result
in poor performance at times. Some better techniques like FedMA [58] and FedPer [2] were presented, by IBM and
Adobe respectively. Some other techniques like Bayesian Nonparametric Federated Learning of Neural Networks
[65], Similarity based Federated Learning – SimFL [29], and Overlap-FedAvg [68] are also proposed in the literature.
5.3. Federated Learning Frameworks
There are various frameworks available for FL. Most of them are open source and provide almost all the func-
tionality required by FL, as previously discussed. Security is achieved using various techniques and the frameworks
support many ML models. Some of the main frameworks are discussed in the Table 2. Other frameworks are FATE
[59] , NVIDIA CLARA [11], Substra [15] and OpenFed [9] that provide support for various different use-cases.
Table 2: Federated Learning Frameworks
S.No Name Developer Description
1.
IBM Feder-
ated Learning
[35,45]
IBM Watson
Project
IBM FL was introduced as a part of the IBM Watson Project that is a part of IBM
Cloud Pak for Data and IBM Cloud Pak for Data as a Service. The framework
provides infrastructure and coordination for federated learning jobs to build
upon existing Deep Learning and Machine Learning algorithms. The frame-
work provides components for both the aggregator and the party (peers/clients).
Training of Neural Networks and Decision Trees is possible using this FL
framework.
2. TensorFlow
Federated [18]Google
It enables the developers to try the existing FL algorithms and experiment with
novel algorithms as well. It has two parts: 1) Federated Learning API, a high
level interface for the developers to use the existing FL for training and evalu-
ation and 2) Federated Learning Core, a lower level interface that can be used
for developing novel FL algorithms.
3.
PySyft +
PyGrid by Py-
Torch [43,42]
OpenMind
PySyft is used together with PyGrid to provide support for FL models. PyGrid
is a peer-to-peer platform where data owners can manage their data as well
as provide and monitor access to data clusters. PySyft allows users to perform
secure deep learning and is built as an extension of PyTorch, TensorFlow and
Keras.
4. Flower [5,4]
Joint project by
Oxford, UCL
and Cambridge
A comprehensive framework that supports large scale FL experiments on Het-
erogenous devices. The main features of this framework are scalability, Support
for heterogenous devices, Realism in results and Privacy preservation.
5. OpenFL [48] Intel Corp
Data Private Collaborative Framework for FL that can be extended to various
ML and DL models and support multiple aggregation algorithms like FedAvg,
FedProx, FedOpt etc.
6. Fed-BioMed
[21]
French Com-
puter Institute-
INRIA
This framework is mostly focused on Biomedical research and provides an FL
analysis framework. It is user friendly and easy to deploy.
7. XayNet [12] Xayn- Berlin It is a masked cross device FL Framework that supports Horizontal and Transfer
FL in both cross-device and cross-silos setting.
6. Applications of Federated Learning
The employment of smart devices and increasing automation in industrial solutions are the main goals of industry
4.0. Due to the decentralized and privacy-preserving nature of FL, it can be used in various fields to provide smart
solutions. An overview of some of the applications:
Healthcare: To keep up with the Industry 4.0 trends, there are research articles that encourage the use of FL in
healthcare and medical informatics due to its privacy-preserving nature. Data-driven medicines and treatments require
patients’ data that is confidential in nature. Thus, FL is the best solution for such cases [49]. FL also addresses many

322 Mashal Khan et al. / Procedia Computer Science 217 (2023) 316–325
Author name /Procedia Computer Science 00 (2022) 000–000 7
statistical and system challenges in the field of Biomedical science by providing access to various clinical, pharmaceu-
tical, and hospital data without compromising the data privacy [61]. Some organizations, like AI Sweden, are using
FL to predict the survival rate of emergency care patients in their project “Federated Mortality Prediction”[53]. An-
other project “FedHealth”[10] uses Federated Transfer Learning to train the model for personalized prediction on the
data collected from wearable health care devices and is tested for Parkinson’s disease. Li et al. [32] implemented and
evaluated FL for Brain Tumor Segmentation on the BraTS data-set and applied differential privacy to guarantee the
alliance with the data-privacy regulations. Fed-BioMed [21] is an open-source front-end framework for FL in medical
science. The practical application of this framework is the analysis of multi-centric brain imaging data. Natural Lan-
guage Processing (NLP) models are used to predict medical illness like depression with the help of FL [3].
Natural Language Processing: FL is also used and has some applications in the domain of Natural Language
Processing (NLP). One of the most famous applications is Google’s GBoard [63] where FL is used to improve query
suggestions in the keyboard for Android phones. Another application is the prediction of the next word in mobile key-
boards [19]. A word-level recurrent neural network with on-device training is used to predict emojis. The technique
shows that Federated Transfer Learning is a better approach for the prediction of emojis as compared to the server
based approach [47]. FedBERT [56] suggests the use of pre-trained BERT models in a federated way, limiting access
to raw data and getting better results.
Computer Vision: AI Sweden, in collaboration with some other organizations, worked on developing a test-bed
for FL, a project called “FedBird”. The project was training an object detection model on two nodes, each of them
having different datasets and sample spaces [54]. Another application of FL in the field of computer vision is a plat-
form known as “FedVision”, used for the detection of objects in videos that are large in size and expensive to transmit
[34].
Autonomous Systems and Internet of Things: Internet of Things (IoT), Cyber-Physical Systems (CPS), and
Edge Computing with AI and ML/DL solutions are the main components of Industry 4.0 processes and with their
increasing use, there is a need to incorporate data privacy and smart processing in the training of the models. To
add trustworthiness into CPS at edge devices and implement smart services, a FL Framework “FengHuoLun”[66]
is used to train machine learning models. Personalized FL models, to deal with the heterogeneity issues in the IoT
environment, also show good results and are tested on Human Activity Recognition Systems utilizing Edge Computing
power of fast computing [60]. The FL-based approach used for Anomaly Detection for IoT devices shows that FL
outperforms the centralized techniques [37].
The use of FL in various IoT and CPS on Edge networks is suggested as trustworthy, smart, and efficient as compared
to the traditional centralized approaches [33,26].
Other Applications: The use of FL is suggested in many application domains such as Recommender Systems
[46], Fraud Detection, FinTech, and Insurance [30], and Communication enhancement [8,57].
7. Research Gaps and Current Challenges
As FL is a new and emerging field, there exist certain challenges and areas of concern. We will now discuss some
of these.
Heterogeneous Devices : One of the major challenges is training on devices having different hardware and soft-
ware structures. Data formatting and storage techniques can also vary from device to device [6].
Non-IID Data and Quality: Handling Non-IID (Independent and Identically Distributed) data is also a challenge
that needs to be addressed. Since raw data is inaccessible in the case of FL, it is difficult to comment on its quality and
authenticity. There are issues of Label Distribution Skew and Attribute Skew in non-IID data. These issues can also
affect the accuracy of the model [69,67].
Communication Bottleneck: There is a communication bottleneck as an additional cost and delay is introduced
while training on various devices and data sharing among them. The model update frequency is also something that
can flood the network, if not managed properly [55].
Unstable Environment and Scalability: The models are trained on various heterogeneous devices and gradients
are sent over the network. The environment is dependent on various factors and can introduce noise or delay. Scala-
bility is also a challenge [33].
Stability in Vertical FL: The Vertical FL is not very mature yet and the mapping of varying feature space across
multiple data sources without compromising User data privacy is still a challenge [38].
8Author name /Procedia Computer Science 00 (2022) 000–000
Resource Constraints: In the cross-device model training, the clients are usually edge devices with minimal re-
sources. The training and communicating of the results can take a little longer due to resource constraints like low
computational power and storage and low communication bandwidth [33].
Possibility of Security Attacks: There are security concerns related in FL in terms of Data and Model Poisoning
[52,22], so proper authentication techniques need to be incorporated at both the data sources and aggregation servers.
Poisoning attack on a large scale can affect the accuracy of the model. To make the system secure, the use of differ-
ential privacy is suggested but that results in a trade-offbetween security and accuracy of the model [38,23]. Replay
and Masquerading attacks are possible even if the weights are encrypted, affecting the performance and accuracy
of the model [36]. Due to the centralized nature of the model, Distributed Denial of Service (DDOS) attack is also
possible on the network and central server. Frequent updates of local models and aggregating it into the global model
can exhaust the central server and flood the network at the same time. Chen et al. [50] evaluated the efficiency of FL in
IoT and concluded that there is a potential vulnerability of Man-in-the-Middle attack. Due to the issue of data leakage
from the gradients, user privacy is at risk even if distributed systems or collaborative training are used [70].
8. Conclusion
Increasing automation and deploying smart devices and processes with access to more data, that can help in in-
creasing productivity, are the main aims of 4th Industrial Revolution. FL is a privacy solution that can enhance the
accuracy of various ML algorithms by shifting the model training to the data sources. It also helps in reducing the load
on the centralized servers thus saving energy and can help in accessing more data by incorporating privacy and data
security at multiple levels. In this paper, the different types of FL models and aggregation algorithms are highlighted
that provide support in the training of ML models. The FL frameworks/platforms discussed in this paper have support
for multiple libraries and ML/DL models. The privacy-preserving nature and trustworthiness of FL make it ideal for
use in different Industry 4.0 application domains, but due to the challenges faced by FL pointed out in this work, it is
not yet fully implemented. We expect that the issues are being addressed with the passage of time and the techniques
used by FL are improving. In the coming years, it is expected that FL will be implemented in most of its envisaged
application areas and that models will become better, more accurate, and more secure.
Acknowledgements
This work was conducted with the financial support of the Science Foundation Ireland Centre for Research Training
in Artificial Intelligence under Grant No. 18/CRT /6223.
References
[1] Alp, D., Acar, E., Zhao, Y., Matas Navarro, R., Mattina, M., Whatmough, P.N., Saligrama, V., 2021. Federated Learning Based on Dynamic
Regularization URL: https://arxiv.org/abs/2111.04263v2, doi:10.48550/arxiv.2111.04263.
[2] Arivazhagan, M.G., Aggarwal, V., Singh, A.K., Choudhary, S., 2019. Federated Learning with Personalization Layers URL: https://arxiv.
org/abs/1912.00818v1, doi:10.48550/arxiv.1912.00818.
[3] Basu, P., Roy, T.S., Naidu, R., Muftuoglu, Z., Singh, S., Mireshghallah, F., 2021. Benchmarking Differential Privacy and Federated Learning
for BERT Models URL: https://arxiv.org/abs/2106.13973v2, doi:10.48550/arxiv.2106.13973.
[4] Beutel, D.J., Topal, T., Mathur, A., Qiu, X., Fernandez-Marques, J., Gao, Y., Sani, L., Li, K.H., Parcollet, T., Porto, P., De Gusm ˜
ao, B., Lane,
N.D., 2020a. Flower: A Friendly Federated Learning Framework. URL: https://flower.dev/.
[5] Beutel, D.J., Topal, T., Mathur, A., Qiu, X., Fernandez-Marques, J., Gao, Y., Sani, L., Li, K.H., Parcollet, T., Porto, P., De Gusm ˜
ao, B.,
Lane, N.D., 2020b. Flower: A Friendly Federated Learning Research Framework URL: https://arxiv.org/abs/2007.14390v5, doi:10.
48550/arxiv.2007.14390.
[6] Bonawitz, K., Eichner, H., Grieskamp, W., Huba, D., Ingerman, A., Ivanov, V., Kiddon, C., Koneˇ
cn´y, J.K., Mazzocchi, S., Mcmahan, H.B.,
Overveldt, T.V., Petrou, D., Ramage, D., Roselander, J., 2019. TOWARDS FEDERATED LEARNING AT SCALE: SYSTEM DESIGN .
[7] Bonawitz, K., Ivanov, V., Kreuter, B., Marcedone, A., Mcmahan, H.B., Patel, S., Ramage, D., Segal, A., Seth, K., . Practical Secure Aggregation
for Privacy-Preserving Machine Learning .
[8] Brendan McMahan Eider Moore Daniel Ramage Seth Hampson Blaise Ag¨
ueraAg, H., Arcas, A., 2017. Communication-Efficient Learning of
Deep Networks from Decentralized Data .
[9] Chen, D., Tan, V., Lu, Z., Hu, J., 2021. OpenFed: A Comprehensive and Versatile Open-Source Federated Learning Framework URL: https:
//arxiv.org/abs/2109.07852v2, doi:10.48550/arxiv.2109.07852.
[10] Chen, Y., Wang, J., Yu, C., Gao, W., Qin, X., . FedHealth: A Federated Transfer Learning Framework for Wearable Healthcare .
[11] Corporation, N., 2019. Federated Learning powered by NVIDIA Clara — NVIDIA Technical Blog. URL: https://developer.nvidia.
com/blog/federated-learning- clara/.
[12] D¨
anschel, W.L., Huth, M., Lundbaek, L.N., . XayNet: Masked Cross-Device Federated Learning Framework .

Mashal Khan et al. / Procedia Computer Science 217 (2023) 316–325 323
Author name /Procedia Computer Science 00 (2022) 000–000 7
statistical and system challenges in the field of Biomedical science by providing access to various clinical, pharmaceu-
tical, and hospital data without compromising the data privacy [61]. Some organizations, like AI Sweden, are using
FL to predict the survival rate of emergency care patients in their project “Federated Mortality Prediction”[53]. An-
other project “FedHealth”[10] uses Federated Transfer Learning to train the model for personalized prediction on the
data collected from wearable health care devices and is tested for Parkinson’s disease. Li et al. [32] implemented and
evaluated FL for Brain Tumor Segmentation on the BraTS data-set and applied differential privacy to guarantee the
alliance with the data-privacy regulations. Fed-BioMed [21] is an open-source front-end framework for FL in medical
science. The practical application of this framework is the analysis of multi-centric brain imaging data. Natural Lan-
guage Processing (NLP) models are used to predict medical illness like depression with the help of FL [3].
Natural Language Processing: FL is also used and has some applications in the domain of Natural Language
Processing (NLP). One of the most famous applications is Google’s GBoard [63] where FL is used to improve query
suggestions in the keyboard for Android phones. Another application is the prediction of the next word in mobile key-
boards [19]. A word-level recurrent neural network with on-device training is used to predict emojis. The technique
shows that Federated Transfer Learning is a better approach for the prediction of emojis as compared to the server
based approach [47]. FedBERT [56] suggests the use of pre-trained BERT models in a federated way, limiting access
to raw data and getting better results.
Computer Vision: AI Sweden, in collaboration with some other organizations, worked on developing a test-bed
for FL, a project called “FedBird”. The project was training an object detection model on two nodes, each of them
having different datasets and sample spaces [54]. Another application of FL in the field of computer vision is a plat-
form known as “FedVision”, used for the detection of objects in videos that are large in size and expensive to transmit
[34].
Autonomous Systems and Internet of Things: Internet of Things (IoT), Cyber-Physical Systems (CPS), and
Edge Computing with AI and ML/DL solutions are the main components of Industry 4.0 processes and with their
increasing use, there is a need to incorporate data privacy and smart processing in the training of the models. To
add trustworthiness into CPS at edge devices and implement smart services, a FL Framework “FengHuoLun”[66]
is used to train machine learning models. Personalized FL models, to deal with the heterogeneity issues in the IoT
environment, also show good results and are tested on Human Activity Recognition Systems utilizing Edge Computing
power of fast computing [60]. The FL-based approach used for Anomaly Detection for IoT devices shows that FL
outperforms the centralized techniques [37].
The use of FL in various IoT and CPS on Edge networks is suggested as trustworthy, smart, and efficient as compared
to the traditional centralized approaches [33,26].
Other Applications: The use of FL is suggested in many application domains such as Recommender Systems
[46], Fraud Detection, FinTech, and Insurance [30], and Communication enhancement [8,57].
7. Research Gaps and Current Challenges
As FL is a new and emerging field, there exist certain challenges and areas of concern. We will now discuss some
of these.
Heterogeneous Devices : One of the major challenges is training on devices having different hardware and soft-
ware structures. Data formatting and storage techniques can also vary from device to device [6].
Non-IID Data and Quality: Handling Non-IID (Independent and Identically Distributed) data is also a challenge
that needs to be addressed. Since raw data is inaccessible in the case of FL, it is difficult to comment on its quality and
authenticity. There are issues of Label Distribution Skew and Attribute Skew in non-IID data. These issues can also
affect the accuracy of the model [69,67].
Communication Bottleneck: There is a communication bottleneck as an additional cost and delay is introduced
while training on various devices and data sharing among them. The model update frequency is also something that
can flood the network, if not managed properly [55].
Unstable Environment and Scalability: The models are trained on various heterogeneous devices and gradients
are sent over the network. The environment is dependent on various factors and can introduce noise or delay. Scala-
bility is also a challenge [33].
Stability in Vertical FL: The Vertical FL is not very mature yet and the mapping of varying feature space across
multiple data sources without compromising User data privacy is still a challenge [38].
8Author name /Procedia Computer Science 00 (2022) 000–000
Resource Constraints: In the cross-device model training, the clients are usually edge devices with minimal re-
sources. The training and communicating of the results can take a little longer due to resource constraints like low
computational power and storage and low communication bandwidth [33].
Possibility of Security Attacks: There are security concerns related in FL in terms of Data and Model Poisoning
[52,22], so proper authentication techniques need to be incorporated at both the data sources and aggregation servers.
Poisoning attack on a large scale can affect the accuracy of the model. To make the system secure, the use of differ-
ential privacy is suggested but that results in a trade-offbetween security and accuracy of the model [38,23]. Replay
and Masquerading attacks are possible even if the weights are encrypted, affecting the performance and accuracy
of the model [36]. Due to the centralized nature of the model, Distributed Denial of Service (DDOS) attack is also
possible on the network and central server. Frequent updates of local models and aggregating it into the global model
can exhaust the central server and flood the network at the same time. Chen et al. [50] evaluated the efficiency of FL in
IoT and concluded that there is a potential vulnerability of Man-in-the-Middle attack. Due to the issue of data leakage
from the gradients, user privacy is at risk even if distributed systems or collaborative training are used [70].
8. Conclusion
Increasing automation and deploying smart devices and processes with access to more data, that can help in in-
creasing productivity, are the main aims of 4th Industrial Revolution. FL is a privacy solution that can enhance the
accuracy of various ML algorithms by shifting the model training to the data sources. It also helps in reducing the load
on the centralized servers thus saving energy and can help in accessing more data by incorporating privacy and data
security at multiple levels. In this paper, the different types of FL models and aggregation algorithms are highlighted
that provide support in the training of ML models. The FL frameworks/platforms discussed in this paper have support
for multiple libraries and ML/DL models. The privacy-preserving nature and trustworthiness of FL make it ideal for
use in different Industry 4.0 application domains, but due to the challenges faced by FL pointed out in this work, it is
not yet fully implemented. We expect that the issues are being addressed with the passage of time and the techniques
used by FL are improving. In the coming years, it is expected that FL will be implemented in most of its envisaged
application areas and that models will become better, more accurate, and more secure.
Acknowledgements
This work was conducted with the financial support of the Science Foundation Ireland Centre for Research Training
in Artificial Intelligence under Grant No. 18/CRT /6223.
References
[1] Alp, D., Acar, E., Zhao, Y., Matas Navarro, R., Mattina, M., Whatmough, P.N., Saligrama, V., 2021. Federated Learning Based on Dynamic
Regularization URL: https://arxiv.org/abs/2111.04263v2, doi:10.48550/arxiv.2111.04263.
[2] Arivazhagan, M.G., Aggarwal, V., Singh, A.K., Choudhary, S., 2019. Federated Learning with Personalization Layers URL: https://arxiv.
org/abs/1912.00818v1, doi:10.48550/arxiv.1912.00818.
[3] Basu, P., Roy, T.S., Naidu, R., Muftuoglu, Z., Singh, S., Mireshghallah, F., 2021. Benchmarking Differential Privacy and Federated Learning
for BERT Models URL: https://arxiv.org/abs/2106.13973v2, doi:10.48550/arxiv.2106.13973.
[4] Beutel, D.J., Topal, T., Mathur, A., Qiu, X., Fernandez-Marques, J., Gao, Y., Sani, L., Li, K.H., Parcollet, T., Porto, P., De Gusm ˜
ao, B., Lane,
N.D., 2020a. Flower: A Friendly Federated Learning Framework. URL: https://flower.dev/.
[5] Beutel, D.J., Topal, T., Mathur, A., Qiu, X., Fernandez-Marques, J., Gao, Y., Sani, L., Li, K.H., Parcollet, T., Porto, P., De Gusm ˜
ao, B.,
Lane, N.D., 2020b. Flower: A Friendly Federated Learning Research Framework URL: https://arxiv.org/abs/2007.14390v5, doi:10.
48550/arxiv.2007.14390.
[6] Bonawitz, K., Eichner, H., Grieskamp, W., Huba, D., Ingerman, A., Ivanov, V., Kiddon, C., Koneˇ
cn´y, J.K., Mazzocchi, S., Mcmahan, H.B.,
Overveldt, T.V., Petrou, D., Ramage, D., Roselander, J., 2019. TOWARDS FEDERATED LEARNING AT SCALE: SYSTEM DESIGN .
[7] Bonawitz, K., Ivanov, V., Kreuter, B., Marcedone, A., Mcmahan, H.B., Patel, S., Ramage, D., Segal, A., Seth, K., . Practical Secure Aggregation
for Privacy-Preserving Machine Learning .
[8] Brendan McMahan Eider Moore Daniel Ramage Seth Hampson Blaise Ag¨
ueraAg, H., Arcas, A., 2017. Communication-Efficient Learning of
Deep Networks from Decentralized Data .
[9] Chen, D., Tan, V., Lu, Z., Hu, J., 2021. OpenFed: A Comprehensive and Versatile Open-Source Federated Learning Framework URL: https:
//arxiv.org/abs/2109.07852v2, doi:10.48550/arxiv.2109.07852.
[10] Chen, Y., Wang, J., Yu, C., Gao, W., Qin, X., . FedHealth: A Federated Transfer Learning Framework for Wearable Healthcare .
[11] Corporation, N., 2019. Federated Learning powered by NVIDIA Clara — NVIDIA Technical Blog. URL: https://developer.nvidia.
com/blog/federated-learning- clara/.
[12] D¨
anschel, W.L., Huth, M., Lundbaek, L.N., . XayNet: Masked Cross-Device Federated Learning Framework .

324 Mashal Khan et al. / Procedia Computer Science 217 (2023) 316–325
Author name /Procedia Computer Science 00 (2022) 000–000 9
[13] Ek, S., Portet, F., Lalanda, P., Vega, G., 2021. A Federated Learning Aggregation Algorithm for Pervasive Computing: Evaluation and Com-
parison. 2021 IEEE International Conference on Pervasive Computing and Communications, PerCom 2021 URL: https://arxiv.org/
abs/2110.10223v1, doi:10.1109/percom50583.2021.9439129.
[14] (EU), R., 2016. General Data Protection Regulation (GDPR) – Official Legal Text. URL: https://gdpr-info.eu/.
[15] Galtier, M.N., Marini, C., 2019. Substra: a framework for privacy-preserving, traceable and collaborative Machine Learning URL: https:
//arxiv.org/abs/1910.11567v1, doi:10.48550/arxiv.1910.11567.
[16] Gooday, A., 2020. Understanding Federated Learning Terminology. URL: https://blog.openmined.org/
federated-learning- types/.
[17] Google, a. Google AI Blog: Federated Learning: Collaborative Machine Learning without Centralized Training Data. URL: https://ai.
googleblog.com/2017/04/federated-learning- collaborative.html.
[18] Google, T., b. TensorFlow Federated. URL: https://www.tensorflow.org/federated.
[19] Hard, A., Rao, K., Mathews, R., Ramaswamy, S., Beaufays, F., Augenstein, S., Eichner, H., Kiddon, C., Ramage, D., 2018. Federated Learning
for Mobile Keyboard Prediction URL: https://arxiv.org/abs/1811.03604v2.
[20] of Health, U.D., (HHS), H.S., . Health Insurance Portability and Accountability Act of 1996 (HIPAA) — CDC. URL: https://www.cdc.
gov/phlp/publications/topic/hipaa.html.
[21] INRIA, . An open-source federated learning framework. - Fed-BioMed.
[22] Jere, M.S., Farnan, T., Koushanfar, F., 2021. A Taxonomy of Attacks on Federated Learning. IEEE Security and Privacy 19, 20–28. doi:10.
1109/MSEC.2020.3039941.
[23] Kairouz, P., McMahan, H.B., Avent, B., Bellet, A., Bennis, M., Bhagoji, A.N., Bonawitz, K., Charles, Z., Cormode, G., 2019. Advances and
Open Problems in Federated Learning. Foundations and Trends in Machine Learning 14, 1–210. URL: https://arxiv.org/abs/1912.
04977v3, doi:10.1561/2200000083.
[24] Kang, J., Xiong, Z., Niyato, D., Yu, H., Liang, Y.C., Kim, D.I., 2019. Incentive design for efficient federated learning in mobile networks: A
contract theory approach. Proceedings - 2019 IEEE VTS Asia Pacific Wireless Communications Symposium, APWCS 2019 doi:10.1109/
VTS-APWCS.2019.8851649.
[25] Kelvin, . Introduction to Federated Learning and Challenges — Towards Data Science. URL: https://towardsdatascience.com/
introduction-to- federated-learning- and-challenges- ea7e02f260ca.
[26] Kholod, I., Yanaki, E., Fomichev, D., Shalugin, E., Novikova, E., Filippov, E., Nordlund, M., 2020. Open-Source Federated Learning Frame-
works for IoT: A Comparative Review and Analysis. Sensors 2021, Vol. 21, Page 167 21, 167. URL: https://www.mdpi.com/1424- 8220/
21/1/167/htmhttps://www.mdpi.com/1424-8220/21/1/167, doi:10.3390/S21010167.
[27] Kim, H., Park, J., Bennis, M., Kim, S.L., 2020. Blockchained on-device federated learning. IEEE Communications Letters 24, 1279–1283.
doi:10.1109/LCOMM.2019.2921755.
[28] Li, L., Fan, Y., Tse, M., Lin, K.Y., 2020a. A review of applications in federated learning. Computers & Industrial Engineering 149, 106854.
doi:10.1016/J.CIE.2020.106854.
[29] Li, Q., Wen, Z., He, B., . Practical Federated Gradient Boosting Decision Trees URL: www.aaai.org.
[30] Li, Q., Wen, Z., Wu, Z., Hu, S., Wang, N., Li, Y., Liu, X., He, B., 2021. A Survey on Federated Learning Systems: Vision, Hype and Reality
for Data Privacy and Protection. IEEE Transactions on Knowledge and Data Engineering doi:10.1109/TKDE.2021.3124599.
[31] Li, T., Sahu, A.K., Talwalkar, A., Smith, V., 2020b. Federated Learning: Challenges, Methods, and Future Directions. IEEE Signal Processing
Magazine 37, 50–60. doi:10.1109/MSP.2020.2975749.
[32] Li, W., Milletar`
ı, F., Xu, D., Rieke, N., Hancox, J., Zhu, W., Baust, M., Cheng, Y., Ourselin, S., Cardoso, M.J., Feng, A., 2019. Privacy-
Preserving Federated Brain Tumour Segmentation. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intel-
ligence and Lecture Notes in Bioinformatics) 11861 LNCS, 133–141. doi:10.1007/978-3- 030-32692- 0{\_}16.
[33] Lim, W.Y.B., Luong, N.C., Hoang, D.T., Jiao, Y., Liang, Y.C., Yang, Q., Niyato, D., Miao, C., 2020. Federated Learning in Mobile Edge
Networks: A Comprehensive Survey. IEEE Communications Surveys and Tutorials 22, 2031–2063. doi:10.1109/COMST.2020.2986024.
[34] Liu, Y., Huang, A., Luo, Y., Huang, H., Liu, Y., Chen, Y., Feng, L., Chen, T., Yu, H., Yang, Q., 2020. FedVision: An Online Visual Object
Detection Platform Powered by Federated Learning. Proceedings of the AAAI Conference on Artificial Intelligence 34, 13172–13179. URL:
https://ojs.aaai.org/index.php/AAAI/article/view/7021, doi:10.1609/AAAI.V34I08.7021.
[35] Ludwig, H., Baracaldo, N., Thomas, G., Zhou, Y., Anwar, A., Rajamoni, S., Ong, Y., Radhakrishnan, J., Verma, A., Sinn, M., Purcell, M.,
Rawat, A., Minh, T., Holohan, N., Chakraborty, S., Whitherspoon, S., Steuer, D., Wynter, L., Hassan, H., Laguna, S., Yurochkin, M., Agarwal,
M., Chuba, E., Abay, A., 2020. IBM Federated Learning: an Enterprise Framework White Paper V0.1 URL: https://arxiv.org/abs/
2007.10987v1, doi:10.48550/arxiv.2007.10987.
[36] Lyu, L., Yu, H., Zhao, J., Yang, Q., 2020. Threats to Federated Learning. Lecture Notes in Computer Science (including subseries Lecture
Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 12500 LNCS, 3–16. doi:10.1007/978-3- 030-63076- 8{\_}1.
[37] Mothukuri, V., Khare, P., Parizi, R.M., Pouriyeh, S., Dehghantanha, A., Srivastava, G., 2022. Federated-Learning-Based Anomaly Detection
for IoT Security Attacks. IEEE Internet of Things Journal 9, 2545–2554. doi:10.1109/JIOT.2021.3077803.
[38] Mothukuri, V., Parizi, R.M., Pouriyeh, S., Huang, Y., Dehghantanha, A., Srivastava, G., 2021. A survey on security and privacy of federated
learning. Future Generation Computer Systems 115, 619–640. doi:10.1016/J.FUTURE.2020.10.007.
[39] Mugunthan, V., Peraire-Bueno, A., Kagal, L., . PrivacyFL: A Simulator for Privacy-Preserving and Secure Federated Learning. Proceedings of
the 29th ACM International Conference on Information & Knowledge Management URL: https://doi.org/10.1145/3340531.3412771,
doi:10.1145/3340531.
[40] Nishio, T., Yonetani, R., 2019. Client Selection for Federated Learning with Heterogeneous Resources in Mobile Edge. IEEE International
Conference on Communications 2019-May. doi:10.1109/ICC.2019.8761315.
[41] Noia, D.I., Tintarev, N., Fatourou, P., Schedl, M., 2022. big trends europe region Recommender Systems under European AI Regulations.

Mashal Khan et al. / Procedia Computer Science 217 (2023) 316–325 325
Author name /Procedia Computer Science 00 (2022) 000–000 9
[13] Ek, S., Portet, F., Lalanda, P., Vega, G., 2021. A Federated Learning Aggregation Algorithm for Pervasive Computing: Evaluation and Com-
parison. 2021 IEEE International Conference on Pervasive Computing and Communications, PerCom 2021 URL: https://arxiv.org/
abs/2110.10223v1, doi:10.1109/percom50583.2021.9439129.
[14] (EU), R., 2016. General Data Protection Regulation (GDPR) – Official Legal Text. URL: https://gdpr-info.eu/.
[15] Galtier, M.N., Marini, C., 2019. Substra: a framework for privacy-preserving, traceable and collaborative Machine Learning URL: https:
//arxiv.org/abs/1910.11567v1, doi:10.48550/arxiv.1910.11567.
[16] Gooday, A., 2020. Understanding Federated Learning Terminology. URL: https://blog.openmined.org/
federated-learning- types/.
[17] Google, a. Google AI Blog: Federated Learning: Collaborative Machine Learning without Centralized Training Data. URL: https://ai.
googleblog.com/2017/04/federated-learning- collaborative.html.
[18] Google, T., b. TensorFlow Federated. URL: https://www.tensorflow.org/federated.
[19] Hard, A., Rao, K., Mathews, R., Ramaswamy, S., Beaufays, F., Augenstein, S., Eichner, H., Kiddon, C., Ramage, D., 2018. Federated Learning
for Mobile Keyboard Prediction URL: https://arxiv.org/abs/1811.03604v2.
[20] of Health, U.D., (HHS), H.S., . Health Insurance Portability and Accountability Act of 1996 (HIPAA) — CDC. URL: https://www.cdc.
gov/phlp/publications/topic/hipaa.html.
[21] INRIA, . An open-source federated learning framework. - Fed-BioMed.
[22] Jere, M.S., Farnan, T., Koushanfar, F., 2021. A Taxonomy of Attacks on Federated Learning. IEEE Security and Privacy 19, 20–28. doi:10.
1109/MSEC.2020.3039941.
[23] Kairouz, P., McMahan, H.B., Avent, B., Bellet, A., Bennis, M., Bhagoji, A.N., Bonawitz, K., Charles, Z., Cormode, G., 2019. Advances and
Open Problems in Federated Learning. Foundations and Trends in Machine Learning 14, 1–210. URL: https://arxiv.org/abs/1912.
04977v3, doi:10.1561/2200000083.
[24] Kang, J., Xiong, Z., Niyato, D., Yu, H., Liang, Y.C., Kim, D.I., 2019. Incentive design for efficient federated learning in mobile networks: A
contract theory approach. Proceedings - 2019 IEEE VTS Asia Pacific Wireless Communications Symposium, APWCS 2019 doi:10.1109/
VTS-APWCS.2019.8851649.
[25] Kelvin, . Introduction to Federated Learning and Challenges — Towards Data Science. URL: https://towardsdatascience.com/
introduction-to- federated-learning- and-challenges- ea7e02f260ca.
[26] Kholod, I., Yanaki, E., Fomichev, D., Shalugin, E., Novikova, E., Filippov, E., Nordlund, M., 2020. Open-Source Federated Learning Frame-
works for IoT: A Comparative Review and Analysis. Sensors 2021, Vol. 21, Page 167 21, 167. URL: https://www.mdpi.com/1424- 8220/
21/1/167/htmhttps://www.mdpi.com/1424-8220/21/1/167, doi:10.3390/S21010167.
[27] Kim, H., Park, J., Bennis, M., Kim, S.L., 2020. Blockchained on-device federated learning. IEEE Communications Letters 24, 1279–1283.
doi:10.1109/LCOMM.2019.2921755.
[28] Li, L., Fan, Y., Tse, M., Lin, K.Y., 2020a. A review of applications in federated learning. Computers & Industrial Engineering 149, 106854.
doi:10.1016/J.CIE.2020.106854.
[29] Li, Q., Wen, Z., He, B., . Practical Federated Gradient Boosting Decision Trees URL: www.aaai.org.
[30] Li, Q., Wen, Z., Wu, Z., Hu, S., Wang, N., Li, Y., Liu, X., He, B., 2021. A Survey on Federated Learning Systems: Vision, Hype and Reality
for Data Privacy and Protection. IEEE Transactions on Knowledge and Data Engineering doi:10.1109/TKDE.2021.3124599.
[31] Li, T., Sahu, A.K., Talwalkar, A., Smith, V., 2020b. Federated Learning: Challenges, Methods, and Future Directions. IEEE Signal Processing
Magazine 37, 50–60. doi:10.1109/MSP.2020.2975749.
[32] Li, W., Milletar`
ı, F., Xu, D., Rieke, N., Hancox, J., Zhu, W., Baust, M., Cheng, Y., Ourselin, S., Cardoso, M.J., Feng, A., 2019. Privacy-
Preserving Federated Brain Tumour Segmentation. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intel-
ligence and Lecture Notes in Bioinformatics) 11861 LNCS, 133–141. doi:10.1007/978-3- 030-32692- 0{\_}16.
[33] Lim, W.Y.B., Luong, N.C., Hoang, D.T., Jiao, Y., Liang, Y.C., Yang, Q., Niyato, D., Miao, C., 2020. Federated Learning in Mobile Edge
Networks: A Comprehensive Survey. IEEE Communications Surveys and Tutorials 22, 2031–2063. doi:10.1109/COMST.2020.2986024.
[34] Liu, Y., Huang, A., Luo, Y., Huang, H., Liu, Y., Chen, Y., Feng, L., Chen, T., Yu, H., Yang, Q., 2020. FedVision: An Online Visual Object
Detection Platform Powered by Federated Learning. Proceedings of the AAAI Conference on Artificial Intelligence 34, 13172–13179. URL:
https://ojs.aaai.org/index.php/AAAI/article/view/7021, doi:10.1609/AAAI.V34I08.7021.
[35] Ludwig, H., Baracaldo, N., Thomas, G., Zhou, Y., Anwar, A., Rajamoni, S., Ong, Y., Radhakrishnan, J., Verma, A., Sinn, M., Purcell, M.,
Rawat, A., Minh, T., Holohan, N., Chakraborty, S., Whitherspoon, S., Steuer, D., Wynter, L., Hassan, H., Laguna, S., Yurochkin, M., Agarwal,
M., Chuba, E., Abay, A., 2020. IBM Federated Learning: an Enterprise Framework White Paper V0.1 URL: https://arxiv.org/abs/
2007.10987v1, doi:10.48550/arxiv.2007.10987.
[36] Lyu, L., Yu, H., Zhao, J., Yang, Q., 2020. Threats to Federated Learning. Lecture Notes in Computer Science (including subseries Lecture
Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 12500 LNCS, 3–16. doi:10.1007/978-3- 030-63076- 8{\_}1.
[37] Mothukuri, V., Khare, P., Parizi, R.M., Pouriyeh, S., Dehghantanha, A., Srivastava, G., 2022. Federated-Learning-Based Anomaly Detection
for IoT Security Attacks. IEEE Internet of Things Journal 9, 2545–2554. doi:10.1109/JIOT.2021.3077803.
[38] Mothukuri, V., Parizi, R.M., Pouriyeh, S., Huang, Y., Dehghantanha, A., Srivastava, G., 2021. A survey on security and privacy of federated
learning. Future Generation Computer Systems 115, 619–640. doi:10.1016/J.FUTURE.2020.10.007.
[39] Mugunthan, V., Peraire-Bueno, A., Kagal, L., . PrivacyFL: A Simulator for Privacy-Preserving and Secure Federated Learning. Proceedings of
the 29th ACM International Conference on Information & Knowledge Management URL: https://doi.org/10.1145/3340531.3412771,
doi:10.1145/3340531.
[40] Nishio, T., Yonetani, R., 2019. Client Selection for Federated Learning with Heterogeneous Resources in Mobile Edge. IEEE International
Conference on Communications 2019-May. doi:10.1109/ICC.2019.8761315.
[41] Noia, D.I., Tintarev, N., Fatourou, P., Schedl, M., 2022. big trends europe region Recommender Systems under European AI Regulations.
10 Author name /Procedia Computer Science 00 (2022) 000–000
COMMUNICATIONS OF THE ACM 65. URL: https://bit.ly/3r3mrH8, doi:10.1145/3512728.
[42] OpenMind, a. GitHub - OpenMined/PyGrid: A Peer-to-peer Platform for Secure, Privacy-preserving, Decentralized Data Science. URL:
https://github.com/OpenMined/PyGrid.
[43] OpenMind, b. PyGrid: A Peer-to-Peer Platform for Private Data Science and Federated Learning. URL: https://blog.openmined.org/
what-is- pygrid-demo/.
[44] Papadopoulos, P., Abramson, W., Hall, A.J., Pitropakis, N., Buchanan, W.J., 2021. Privacy and Trust Redefined in Federated Machine Learning.
Machine Learning and Knowledge Extraction 2021, Vol. 3, Pages 333-356 3, 333–356. URL: https://www.mdpi.com/2504-4990/3/2/
17/htmhttps://www.mdpi.com/2504-4990/3/2/17, doi:10.3390/MAKE3020017.
[45] Project, I.W., . IBM Federated Learning. URL: https://ibmfl.mybluemix.net/github.
[46] Qi, T., Wu, F., Wu, C., Huang, Y., Xie, X., 2020. Privacy-Preserving News Recommendation Model Learning. Findings of the Association for
Computational Linguistics Findings of ACL: EMNLP 2020 , 1423–1432URL: https://arxiv.org/abs/2003.09592v3, doi:10.48550/
arxiv.2003.09592.
[47] Ramaswamy, S., Mathews, R., Rao, K., Beaufays, F., 2019. Federated Learning for Emoji Prediction in a Mobile Keyboard URL: http:
//arxiv.org/abs/1906.04329.
[48] Reina, G.A., Gruzdev, A., Foley, P., Perepelkina, O., Sharma, M., Davidyuk, I., Trushkin, I., Radionov, M., Mokrov, A., Agapov, D., Martin, J.,
Edwards, B., Sheller, M.J., Pati, S., Moorthy, P.N., Wang, S.h., Shah, P., Bakas, S., 2021. OpenFL: An open-source framework for Federated
Learning URL: https://arxiv.org/abs/2105.06413v1, doi:10.48550/arxiv.2105.06413.
[49] Rieke, N., Hancox, J., Li, W., Milletar`
ı, F., Roth, H.R., Albarqouni, S., Bakas, S., Galtier, M.N., Landman, B.A., Maier-Hein, K., Ourselin, S.,
Sheller, M., Summers, R.M., Trask, A., Xu, D., Baust, M., Cardoso, M.J., . PERSPECTIVE The future of digital health with federated learning
URL: https://doi.org/10.1038/s41746-020- 00323-1, doi:10.1038/s41746- 020-00323- 1.
[50] Shen, C., Xue, W., 2022. An Experiment Study on Federated Learning Testbed. Lecture Notes in Networks and Systems 286, 209–217.
doi:10.1007/978-981- 16-4016- 2{\_}20.
[51] Shokri, R., Shmatikov, V., . Privacy-Preserving Deep Learning. Proceedings of the 22nd ACM SIGSAC Conference on Computer and
Communications Security URL: http://dx.doi.org/10.1145/2810103.2813687., doi:10.1145/2810103.
[52] Song, M., Wang, Z., Zhang, Z., Song, Y., Wang, Q., Ren, J., Qi, H., 2020. Analyzing User-Level Privacy Attack against Federated Learning.
IEEE Journal on Selected Areas in Communications 38, 2430–2444. doi:10.1109/JSAC.2020.3000372.
[53] Sweden, A., . POCs within the testbed. URL: https://www.ai.se/en/node/84340.
[54] Sweden, A., 2021. Federated Learning Testbed — AI Sweden. URL: https://www.ai.se/en/node/81535/
federated-learning- testbed.
[55] Tan, J., Liang, Y.C., Luong, N.C., Niyato, D., 2021. Toward Smart Security Enhancement of Federated Learning Networks. IEEE Network
35, 340–347. doi:10.1109/MNET.011.2000379.
[56] TianYuanyishu, WanYao, LyuLingjuan, YaoDezhong, JinHai, SunLichao, 2021. FedBERT: When Federated Learning Meets Pre-Training.
ACM Transactions on Intelligent Systems and Technology (TIST) URL: https://dl.acm.org/doi/abs/10.1145/3510033, doi:10.
1145/3510033.
[57] Wang, H., Qu, Z., Guo, S., Gao, X., Li, R., Ye, B., 2020a. Intermittent Pulling with Local Compensation for Communication-Efficient
Distributed Learning. IEEE Transactions on Emerging Topics in Computing doi:10.1109/TETC.2020.3043300.
[58] Wang, H., Yurochkin, M., Sun, Y., Papailiopoulos, D., Khazaeni, Y., 2020b. Federated Learning with Matched Averaging URL: https:
//arxiv.org/abs/2002.06440v1, doi:10.48550/arxiv.2002.06440.
[59] WeBank, C., . Fate. URL: https://fate.fedai.org/.
[60] Wu, Q., He, K., Chen, X., 2020. Personalized Federated Learning for Intelligent IoT Applications: A Cloud-Edge based Framework. IEEE
Computer Graphics and Applications doi:10.1109/OJCS.2020.2993259.
[61] Xu, J., Glicksberg, B.S., Su, C., Walker, P., Bian, J., Wang, F., 2021. Federated Learning for Healthcare Informatics. Journal of Healthcare
Informatics Research 5. doi:10.1007/S41666-020- 00082-4.
[62] Yang, Q., Liu, Y., Cheng, Y., Kang, Y., Chen, T., Yu, H., 2019. Federated Learning. Synthesis Lectures on Artificial Intelligence and Machine
Learning 13, 1–207. doi:10.2200/S00960ED2V01Y201910AIM043.
[63] Yang, T., Andrew, G., Eichner, H., Sun, H., Li, W., Kong, N., Ramage, D., Beaufays, F., 2018. Applied Federated Learning: Improving Google
Keyboard Query Suggestions URL: https://arxiv.org/abs/1812.02903v1, doi:10.48550/arxiv.1812.02903.
[64] Yoshida, N., Nishio, T., Morikura, M., Yamamoto, K., 2020. MAB-based Client Selection for Federated Learning with Uncertain Resources
in Mobile Networks. 2020 IEEE Globecom Workshops, GC Wkshps 2020 - Proceedings doi:10.1109/GCWKSHPS50303.2020.9367421.
[65] Yurochkin, M., Agarwal, M., Ghosh, S., Greenewald, K., Hoang, T.N., Khazaeni, Y., . Bayesian Nonparametric Federated Learning of Neural
Networks URL: https://github.com/IBM/.
[66] Zhang, C., Liu, X., Zheng, X., Li, R., Liu, H., 2020. FengHuoLun: a federated learning based edge computing platform for cyber-physical sys-
tems. 2020 IEEE International Conference on Pervasive Computing and Communications Workshops, PerCom Workshops 2020 URL: https:
//researchers.mq.edu.au/en/publications/fenghuolun-a- federated-learning- based-edge- computing-platform-for,
doi:10.1109/PERCOMWORKSHOPS48775.2020.9156259.
[67] Zhao, Y., Li, M., Lai, L., Suda, N., Civin, D., Chandra, V., . Federated Learning with Non-IID Data. Technical Report.
[68] Zhou, Y., Ye, Q., Lv, J., . Communication-Efficient Federated Learning with Compensated Overlap-FedAvg URL: https://github.com/
Soptq/Overlap-FedAvg.
[69] Zhu, H., Xu, J., Liu, S., Jin, Y., 2021. Federated learning on non-IID data: A survey. Neurocomputing 465, 371–390. doi:10.1016/J.
NEUCOM.2021.07.098.
[70] Zhu, L., Liu, Z., Han, S., 2019. Deep Leakage from Gradients. URL: https://proceedings.neurips.cc/paper/2019/hash/
60a6c4002cc7b29142def8871531281a-Abstract.html.