Empirical Analysis of Federated Learning Algorithms: A Federated
Research Infrastructure Use Case
Harshit Gupta
1
, Abhishek Verma
1
, O. P. Vyas
1
, Marco Garofalo
2,3
, Giuseppe Tricomi
2,3
,
Francesco Longo
2,3
, Giovanni Merlino
2,3
and Antonio Puliafito
2,3
1
IIIT Allahabad, India
2
University of Messina, Italy
3
CINI, Italy
marco.garofalo@studenti.unime.it, {gtricomi, flongo, gmerlino, apuliafito}@unime.it
Keywords:
Federated Learning, Federated Research Infrastructure, FedAvg, FedProx, Stragglers, Statistical
Heterogeneity.
Abstract:
Research Infrastructures provide resources and services for communities of researchers at large to conduct
their experiments and foster innovation. Moreover, these can be used beyond research, e.g., for education or
public service. The SLICES consortium is chartered to provide a fully programmable, distributed, virtual-
ized, remotely accessible, European-wide, federated research infrastructure, providing advanced computing,
storage, and networking capabilities, including interconnection by dedicated high-speed links. It will sup-
port large-scale, experimental research across various scientific domains. Data processing, in general, and
especially Machine Learning, are of great interest to the potential audience of SLICES. According to these
premises, this work aims to exploit such a peculiar Research Infrastructure and its Cloud-oriented development
and deployment facilities to investigate Federated Learning (FL) approaches; in particular, here we evaluate
the performance of two FL aggregation algorithms, i.e., FedAvg and FedProx, in settings, characterized by
system heterogeneity, and statistical heterogeneity, that represent plausible, and possibly common, scenarios
in forthcoming facilities, such as those mentioned above, community-oriented, shared Research Infrastruc-
tures. We have observed that the FedProx algorithm outperforms the FedAvg algorithm in such settings.
1 INTRODUCTION
Federated Learning represents a fundamental
paradigm in Deep Learning models and a peculiar
deployment model across Research Infrastruc-
tures. The Scientific Large Scale Infrastructure for
Computing/Communication Experimental Studies
consortium, commonly indicated as SLICES, intends
to provide research communities with a rich environ-
ment to conduct experiments and foster innovation,
a fully programmable and virtualized, distributed
and remotely accessible one, providing advanced
computing, storage and networking capabilities, in-
cluding inter-site connectivity by means of dedicated
high-speed links, i.e., granted by consortium-level
agreements with G
´
EANT, which is a collaboration
among European National Research and Educa-
tion Networks (NRENs). Under G
´
EANT, NRENs
together deliver an information ecosystem of infras-
tructure and services to advance research, education,
and innovation on a global scale. SLICES aims
for a common research infrastructure that combines
diverse technologies and services with geographically
distributed research ambitions (SLICES-RI, 2021b).
Machine Learning is growing and becoming es-
sential as more and more researchers and industries
adopt it, and Federated Learning is a groundbreaking
approach to train models as near as possible to indi-
vidual user data (Li et al., 2019). Federated Learn-
ing enhances traditional machine learning approaches
with the additional (built-in) property of protecting
(training) data in terms of privacy. This strategy has
been implemented by several incumbent tech firms,
including Google, the inventor of federated learning
(Dhada et al., 2020), and Apple, a significant com-
petitor in the field with its virtual assistant technology.
Figure 1 represents how FL may work as a typ-
ical collaboration pattern for similar research activ-
ities, which can be accomplished remotely, without
sharing critical data, over the research infrastructure.
324
Gupta, H., Verma, A., Vyas, O., Garofalo, M., Tricomi, G., Longo, F., Merlino, G. and Puliafito, A.
Empirical Analysis of Federated Learning Algorithms: A Federated Research Infrastructure Use Case.
DOI: 10.5220/0012037600003488
In Proceedings of the 13th International Conference on Cloud Computing and Services Science (CLOSER 2023), pages 324-331
ISBN: 978-989-758-650-7; ISSN: 2184-5042
Copyright
c
2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)
In this diagram, the layer at the top depicts Feder-
ated Learning, comprising geographically distributed
users at the lower layer via a network backbone as the
middle layer.
Figure 1: Federated Learning Comprised Research Infras-
tructure.
The research activities being performed at geo-
graphically disconnected areas for specific purposes
may be brought together, and the (e.g., training) per-
formance or (e.g., inference) accuracy achieved may
be higher with the backing of Federated Learning
methodologies within a research infrastructure.
As FL applications are expanding, getting into
the area of Healthcare by protecting patient’s medical
data, thus keeping confidentially(Gope et al., 2021),
as well as the Insurance sector to predict the risk as-
sociated with customers(Gupta et al., 2022), or, e.g.,
prediction of keyboard(Hard et al., 2018) input, it now
represents a quickly growing research area. A re-
search infrastructure with federated learning as a ser-
vice offered to potential users gives a common thread
to researchers who benefit from data privacy, per-
formance enhancement, and reduced communication
overhead.
Counterbalancing its intrinsic capabilities, FL suf-
fers from the issue of heterogeneity in terms of the de-
vice (system heterogeneity) and data (statistical het-
erogeneity). In terms of the former, in general FL
may engage nodes with uneven computation and/or
storage capabilities to participate in the training pro-
cess (De Vita and Bruneo, 2020). Also, the distribu-
tion of such private data among all the participating
clients can be Independent and Identically Distributed
(IID) or Non-Independent and Identically Distributed
(NIID) (Long et al. 2021). Figure 2 provides an
example of data distribution in IID and NIID envi-
ronments. As can be observed in IID data, the data
among the two clients are similar, while in NIID data,
both clients have different data classes. The NIID
data produces the issue of statistical heterogeneity(Li
et al., 2018).
In an attempt to handle heterogeneity and tackle
high communication costs, differential privacy (Wei
Figure 2: Distribution of Data in IID and NIID Environ-
ment.
et al., 2020) for security, optimization methods that
allow for local updating and low participation are a
popular approach for federated learning(Gad, 2020).
In particular, FedAvg (McMahan et al., 2017) is an
iterative method that has emerged as the de facto op-
timization method in federated settings.
The FedAvg algorithm performs fixed E epochs
of Stochastic Gradient Descent (SGD) on K devices,
where E is a small constant and K is a small fraction
of the total devices in the network. In the presence of
stragglers, i.e., devices which perform slowly, the Fe-
dAvg drops them if they are unable to perform E local
epochs in due time. On the other hand, with dissimi-
lar (heterogeneous) local objective F
k
, a large number
of local epochs may lead each device towards the op-
tima of its local objective as opposed to the global ob-
jective, potentially causing the method to diverge and
thus lacking convergence guarantees to characterize
its behaviour.
It has been seen that due to the presence of such
heterogeneity, the FL performance degrades while
working with the FedAVG algorithm. To minimize
the effect of system and statistical heterogeneity in
the FL setup, authors have proposed a FedProx algo-
rithm(Li et al., 2018) which is an optimized form of
the FedAvg algorithm with the additional capacity to
deal with stragglers and NIID data.
As FedAvg drops the devices which perform par-
tial model updates, i.e., unable to perform fixed E
epochs, it only considers the devices which perform
complete E epochs. But FedProx does not drop out
such slower devices, instead taking partial model up-
dates from such devices as well.
Moreover, to limit the impact of varying local
updates, a proximal term is added to the local sub-
problem. In particular, instead of just minimizing the
local function F
k
(.), device k uses its local solver of
choice to approximately minimize the following ob-
jective h
k
:
Empirical Analysis of Federated Learning Algorithms: A Federated Research Infrastructure Use Case
325
min
w
h
k
w; w
t
= F
k
(w) +
µ
2
w w
t
2
(1)
where
µ
2
w w
t
2
is the proximal term while w and
w
t
are weight associated with local and global model
respectively. µ is a hyperparameter that controls the
strength of the proximal term.
Adding a proximal term to the objective helps to
improve the stability of the method and improve the
overall accuracy of federated learning in heteroge-
neous networks, significantly improving the absolute
testing accuracy on average in highly heterogeneous
settings.
Hence the work focuses on evaluating the per-
formance of the FL algorithms, i.e. FedAvg and
FedProx, in different heterogeneous environments
with different datasets and models. Empirically we
demonstrated that the performance of FedProx is bet-
ter than FedAvg in the case of NIID data. This result
is obtained through the exploitation of SLICES Re-
search Infrastructure to provide some insight on how
exploiting it to support the users in organizing and
planning their experiments.
The paper is organized as follows: Section II gives
us an overview of the current literature about FL,
Section III describes the methodology and simula-
tion, Section IV presents the results and their anal-
ysis, and, at last, Section V outlines our conclusions
and planned future scope of the work.
2 LITERATURE SURVEY
Federated Learning (FL) is a machine learning set-
ting in which the user’s critical data is kept locally to
the user where a local model is trained. In contrast
to other ways to train deep learning models in which
user input is pooled and processed at a centralized
site, as described in Figure 3, the job of the central
server in federated learning is to aggregate training
done on the participating devices with their local data.
In FL, the model updates are sent in rounds in which
devices train the model using a local approach such
as Stochastic Gradient Descent, each device sending
the (current, locally-trained) model back to the server.
Then the server aggregates the model together. The
model can then be distributed again. Figure 3 repre-
sents the process flow in FL architecture.
In terms of privacy, federated learning has evident
advantages. It solves the data privacy problems raised
by previous machine learning systems by allowing
data to stay decentralized, making it more difficult for
attackers to access information from any user. This is
Figure 3: Federated Learning Architecture.
because the data never leaves the devices, forcing at-
tackers to circumvent device security rather than seek-
ing to steal data, while it is more vulnerable in tran-
sit(Nishio and Yonetani, 2019). Despite this advan-
tage, there are numerous difficulties and opportunities
for advancement in federated learning, which we dis-
cuss further in later sections.
In (Li et al., 2018), the authors have proposed the
FedProx algorithm by adding a proximal term to the
FedAvg algorithm to deal with statistical and system
heterogeneity. Statistical heterogeneity is the main
problem and challenge in federated learning, as it has
been observed that its behaviour shows a significant
change when data distribution among the participat-
ing devices is NIID. If the data points of the individ-
ual devices diverge from their point of source, then
the performance of the federated learning slightly de-
creases. It also increases model complexity.
Li et al. (Pan and Yang, 2010) mention some so-
lutions for training in heterogeneous environments,
such as local learning and aggregating the local up-
dates from the local devices sent to the centralized
server. This should continue until we hit some good
performance metrics.
The SLICES research infrastructure focuses on
the Internet of Things and Internet of Services,
cloud/edge/fog computing, artificial intelligence, and
many more. (SLICES-SC, 2022) along with dis-
tributed systems (SLICES-RI, 2021a) research, which
mandates the need for privacy-preserving artificial
intelligence methodologies in order to comply with
the EU General Data Protection Regulation (GDPR,
2022) directives. This brings forth the role of Fed-
erated Learning in the context of SLICES, i.e., in a
privacy-preserving, heterogeneous environment (e.g.,
in SLICES, data and/or code can be made inaccessi-
ble to other research infrastructure users, if so desired,
e.g., due to some entity/company policy or other spe-
cific requirements.
The natural heterogeneity of sites participat-
ing, i.e., federated to assemble, the whole, EU-
wide SLICES research infrastructure, and privacy-
preserving facilities, lends itself naturally to experi-
mentation and validation of Federated Learning ap-
proaches.
CLOSER 2023 - 13th International Conference on Cloud Computing and Services Science
326
The studies performed on the heterogeneity in FL
have captured various factors that negatively influ-
ence the global model’s performance. The computa-
tion capability of participating devices makes the en-
tire system slow and degrades global performance. At
the same time, the NIID data among the participating
clients gives statistical heterogeneity to the FL setup,
which restricts the system from convergence by de-
grading the overall performance. In FedProx, a proxi-
mal term was found to be an essential step in boosting
the performance of the existing model while alleviat-
ing the issue of stragglers compared to other state-
of-the-art architectures. One of the most significant
benefits of FedProx is that it guarantees convergence.
Evidence exists that FedAvg is an excellent algorithm
for training the data over local devices, but the conver-
gence problem cannot decipher the features entirely.
An additional proximal term would help in boosting
the accuracy of models in many existing works.
3 METHODOLOGY &
SIMULATION
This work conducted an empirical analysis of the Fe-
dAvg and the FedProx algorithms in different IID and
NIID data environments by leveraging the SLICES
research infrastructure, at the same time, publish-
ing analyzed methods to the catalogue for other re-
searchers, possibly to consume.
3.1 Datasets and Models
The simulation is performed on different datasets with
their IID and NIID properties. The MNIST with IID
and NIID data and CIFAR10 dataset with IID and
NIID data are used in the proposed work.
In the case of IID data, the MNIST dataset con-
tains 70000 samples of handwritten digits, with 60000
samples for training purposes and 10000 samples for
testing purposes. With NIID data, the data are dis-
tributed differently, as shown in Figure 4. In this case,
each client has digits that are not matched to other
clients, making it independent non-identical data. On
the other hand, the CIFAR10 dataset contains 60000
samples with 50000 images for training and 10000 for
testing. The images(Dhada et al., 2020) are obtained
and preprocessed with operations including segmen-
tation, cropping and other processes before feeding it
to the classification model.
In preprocessing the images, the 28x28 MNIST
images and 32x32 CIFAR10 images are flattened into
matrices. Before that, its dimension is changed to 3D,
i.e., 28x28 and one is for the grayscale channel in the
case of MNIST, and 32x32 with 3 RGB channels for
CIFAR10.
The training of the models is done in a federated
approach, where every client keeps its local model,
which synchronizes with the global model available
in the cloud using the model updates. In this work,
two models, i.e. deep Convolutional Neural Network
(CNN) (Albawi et al., 2017) and Multilayer Percep-
tron (MLP) (Taud and Mas, 2018) models, are used
for training in different environments with different
data distribution and presence of stragglers.
Figure 4: Clients with NIID.
3.2 Simulation Methodology
To analyze the FedAvg & FedProx algorithms, the
simulation is performed with 0%, 50% and 90% strag-
glers and proximal term value µ = 0.3. The simula-
tion consists of 100 participating clients, a 0.1 learn-
ing rate and maximum 80 numbers of communication
rounds between clients and servers. The following
combinations of models and datasets are used in the
simulation for 0%, 50% and 90% stragglers and prox-
imal term value & µ = 0.3.
1. FedAvg with CNN model on IID & NIID data
properties of MNIST and CIFAR10 datasets.
2. FedAvg with MLP model on IID & NIID data
properties of MNIST and CIFAR10 dataset.
3. FedProx with CNN model on IID & NIID data
properties of MNIST and CIFAR10 dataset.
4. FedProx with MLP model on IID & NIID data
properties of MNIST and CIFAR10 dataset.
3.3 Experiments Setup
The simulation is realized on the platform “SoBig-
Data”
1
(Giannotti et al. 2018)(Cresci et al., 2019),
one of the gateway portals towards the federated
SLICES research infrastructure, as well as the name-
sake of the Italian (country-wide) site of SLICES, by
configuring the experiments via a method published
and made available on the infrastructure. The pub-
lished method performs the simulation requested ac-
cording to the parameters used during the experiment
execution and produces the aimed results.
1
SoBigData e-infrastructure portal:
https://sobigdata.d4science.org
Empirical Analysis of Federated Learning Algorithms: A Federated Research Infrastructure Use Case
327
The research infrastructure, to cope with this goal,
provides two valuable tools for tests preparation:
1. The Method Development page (see Figure 5) is
a dedicated environment based on JupyterHub,
where through a notebook or console, the devel-
oper can write, develop, and interactively test his
code in a sandbox provided with useful and ready-
to-use libraries;
2. The pair of functions: Method Importer and
Method Engine. Respectively, the former is used
to upload and publish a method, and the latter is
used to run the experiments according to the sim-
ulation parameters.
Figure 5: SoBigData platform: Method Development
Launcher.
After the development phase, the method just im-
plemented gets published (either as a black box, or
the code shared), when deemed worthwhile, and its
sharing is consistent with policies to foster code reuse
across different experiments. The Method Importer
generates a project for each method published and
needs the definition of several elements such as in-
put/output, code interpreter, general information, and
so on. Figure 6 shows a view of the Compari-
son FedAVG Fedprox method. Finally, the experi-
ments can be performed by launching the method
via the opportune invocation panel (e.g., the function
named Execute an experiment), and as it is shown in
Figure 7, it provides a selection panel in which select
the method to be executed according to the category
and the description available.
4 RESULT & ANALYSIS
To analyze the FL aggregation algorithms, FedAvg
and FedProx, simulations of different scenarios have
been performed by executing the published methods
with suitable inputs to dictate the expected method’s
behaviour and corresponding results stored for further
analysis. A careful value assignment to µ is important
Figure 6: SoBigData platform: Method Importer view.
Figure 7: SoBigData platform: Method Invoker view.
in the FedProx algorithm. So for the entire simula-
tion, the value of µ is taken as 0.3.
Table 1 shows the accuracy obtained when the
probability of stragglers is 0%, meaning there is no
straggler in the entire FL setup. When the dataset is
MNIST, and the model is a CNN, the accuracy of Fed-
Prox is 90.91%, but it is 81.90% in the case of the Fe-
dAvg algorithm. So It can be deduced that in the case
of NIID data, the FedProx algorithm performed bet-
ter than the FedAvg algorithm. Also, the performance
of the FedProx algorithm is better in the case of IID
data as compared to NIID data, which is represented
in figures 8 and 9.
Figures 10 and 11 represent the comparative per-
formance of FedAvg and FedProx for all the possible
combinations of simulations performed in this work.
Here in the case of any dataset or any model, the per-
formance of the FedProx algorithm is better for NIID
data than the FedAvg algorithm. While in the case
of only IID data, the performance difference between
FedAvg and FedProx algorithms is minimal. So in
any setup where the distribution is IID, the FedAvg
can be preferred, but when the distribution becomes
NIID, adopting the FedProx algorithm would show its
benefits. So it can be confirmed that the FedProx can
deal with the issue of statistical heterogeneity in FL.
Figure 8: Accuracy Graph for FedProx in IID data.
CLOSER 2023 - 13th International Conference on Cloud Computing and Services Science
328
Table 1: Performance comparison for FedAvg and Fed-
Prox for 100 Clients , 0% stragglers, µ=0.3 and learning
rate=0.01.
Dataset Distribution Algorithm Model Accuracy(%)
MNIST IID FedAvg CNN 93
MNIST IID FedProx CNN 92.66
MNIST Non-IID FedAvg CNN 81.90
MNIST Non-IID FedProx CNN 90.91
MNIST IID FedAvg MLP 90.28
MNIST IID FedProx MLP 89.87
MNIST Non-IID FedAvg MLP 76.09
MNIST Non-IID FedProx MLP 82.33
CIFAR IID FedAvg CNN 83.03
CIFAR IID FedProx CNN 81.75
CIFAR Non-IID FedAvg CNN 77.28
CIFAR Non-IID FedProx CNN 79.17
CIFAR IID FedAvg MLP 76.23
CIFAR IID FedProx MLP 75.97
CIFAR Non-IID FedAvg MLP 69.14
CIFAR Non-IID FedProx MLP 73.56
Figure 9: Accuracy Graph for FedProx in NIID data.
The results obtained with 50% probability for the
presence of stragglers, 100 participating clients, and
µ=0.3 is shown in Table 2. The accuracy of the Fed-
Prox algorithm in the NIID data case is always greater
than the FedAvg algorithm. So it can be concluded
that the performance of the FedProx algorithm is bet-
ter in the case of NIID data, as compared to the Fe-
dAvg algorithm. It is also observed that with the
MNIST dataset, the FedProx works better than Fe-
dAvg in the case of IID data but, on the other hand, in
the case of the CIFAR10 dataset, FedAvg works better
than FedProx in IID data.
The performance of FedProx with MNIST IID
distribution is better than CIFAR10 IID distribution
which can be seen in Figures 12 and 13. Hence, it
can be inferred that in NIID data with 50% stragglers,
the performance of FedProx is better than FedAvg.
However, in the case of IID data with 50% stragglers,
the performance of the FedProx algorithm mostly de-
pends on the type and properties of the dataset be-
ing used. Thus, by these results, it can be deduced
that FedProx can deal both with system heterogene-
ity (i.e., characterized by the occurrence of stragglers)
and statistical heterogeneity (i.e., dataset distributions
and their properties).
Table 3 shows the accuracy obtained when the per-
Figure 10: Performance Comparison of FedAvg and Fed-
Prox with 0% stragglers and µ=0.3 using CIFAR10 dataset.
Figure 11: Performance Comparison of FedAvg and Fed-
Prox with 0% stragglers and µ=0.3 using MNIST dataset.
centage of occurrence of stragglers is 90%. In this
scenario, it has been observed that in both cases, IID
and NIID data, the performance of the FedProx al-
gorithm is better than FedAvg. Figures 14 and 15
show the accuracy at different rounds for the FedAvg
and FedProx algorithms with IID data. Similarly, Fig-
ures 16 and 17 represent the accuracy of the FedAvg
and FedProx algorithms at different rounds with NIID
distribution. Looking at these simulation results, it
can be deduced that FedProx performs better in in-
tense heterogeneous environments as compared to the
FedAvg algorithm.
Putting together all the previous results, it can be
deduced that the FedProx algorithm is always pre-
ferred to the FedAvg algorithm, especially when deal-
ing with both system and statistical heterogeneity.
In the SLICES research infrastructure where hetero-
geneous sites (e.g., featuring the availability of re-
Table 2: Performance comparison for FedAvg and Fed-
Prox for 100 Clients , 50% stragglers, µ=0.3 and learning
rate=0.01.
Dataset Distribution Algorithm Model Accuracy(%)
MNIST IID FedAvg CNN 97.34
MNIST IID FedProx CNN 99.08
MNIST Non-IID FedAvg CNN 89.24
MNIST Non-IID FedProx CNN 96.57
MNIST IID FedAvg MLP 91.55
MNIST IID FedProx MLP 92.46
MNIST Non-IID FedAvg MLP 85.72
MNIST Non-IID FedProx MLP 94.07
CIFAR IID FedAvg CNN 64.09
CIFAR IID FedProx CNN 52.89
CIFAR Non-IID FedAvg CNN 84.47
CIFAR Non-IID FedProx CNN 94.59
CIFAR IID FedAvg MLP 55.18
CIFAR IID FedProx MLP 51.48
CIFAR Non-IID FedAvg MLP 77.28
CIFAR Non-IID FedProx MLP 93.30
Empirical Analysis of Federated Learning Algorithms: A Federated Research Infrastructure Use Case
329
Figure 12: Performance of FedProx on CIFAR10 IID data
with 50% Stragglers.
Figure 13: Performance of FedProx on MNIST IID data
with 50% Stragglers.
Figure 14: Performance of FedAvg on MNIST IID data
with 90% Stragglers.
Figure 15: Performance of FedProx on MNIST IID data
with 90% Stragglers.
sources with wide-ranging computational and/or stor-
age capabilities) establish together a federation, the
benefits of the FedProx algorithm to perform dis-
tributed privacy-preserving training without signifi-
cant performance degradation (e.g., due to the in-
evitable presence of stragglers) can be highlighted
distinctly.
Figure 16: Performance of FedAvg on MNIST NIID data
with 90% Stragglers.
Figure 17: Performance of FedProx on MNIST NIID data
with 90% Stragglers.
Table 3: Performance comparison for FedAvg and Fed-
Prox for 100 Clients , 90% stragglers, µ=0.3 and learning
rate=0.01.
Dataset Distribution Algorithm Model Accuracy(%)
MNIST IID FedAvg CNN 98.37
MNIST IID FedProx CNN 99.08
MNIST Non-IID FedAvg CNN 92.08
MNIST Non-IID FedProx CNN 91.05
MNIST IID FedAvg MLP 82.72
MNIST IID FedProx MLP 93.68
MNIST Non-IID FedAvg MLP 89.09
MNIST Non-IID FedProx MLP 94.19
CIFAR IID FedAvg CNN 79.52
CIFAR IID FedProx CNN 87.43
CIFAR Non-IID FedAvg CNN 83.12
CIFAR Non-IID FedProx CNN 90.32
CIFAR IID FedAvg MLP 73.39
CIFAR IID FedProx MLP 75.27
CIFAR Non-IID FedAvg MLP 71.42
CIFAR Non-IID FedProx MLP 76.68
5 CONCLUSIONS AND FUTURE
SCOPE
The proposed work performed an empirical analysis
of the FedAvg and FedProx algorithms to evaluate
their behaviour in the presence of system and statis-
tical data heterogeneity. The simulation is performed
by varying the number of stragglers to check the per-
formance in the presence of IID and NIID data dis-
tributions. The simulation of the FedAvg and Fed-
Prox algorithms performed on MNIST and CIFAR10
in IID and NIID scenarios shows that FedProx han-
CLOSER 2023 - 13th International Conference on Cloud Computing and Services Science
330
dles NIID data much better than FedAvg. Also, the
CNN model performs better than MLP in this simula-
tion arrangement. It is found that FedProx with strag-
gling clients outperformed FedAvg, likely due to the
inherent randomness of the client’s selection.
The work opens the area for future research where
such algorithms could be analysed in more complex
environments with multiple probabilities of stragglers
and different data distributions. The results obtained
in this work show promise about the suitability of
Federated Learning approaches in the SLICES re-
search infrastructure, where heterogeneous devices
with their private data distribution may participate
in a (common) experiment to achieve the benefits
of programmable environments with many participat-
ing clients without relinquishing privacy concerns and
unique site requirements.
ACKNOWLEDGMENT
This work was partially supported by the Hori-
zon Europe “Scientific Large-Scale Infrastructure for
Computing/Communication Experimental Studies-
preparation project” (SLICES-PP), under grant
101079774.
REFERENCES
Saad Albawi, Tareq Abed Mohammed, and Saad Al-Zawi.
“Understanding of a convolutional neural network”.
In: 2017 International Conference on Engineering and
Technology (ICET). 2017,
Cresci, S., Minutoli, S., Nizzoli, L., Tardelli, S., Tesconi,
M. (2019). Enriching Digital Libraries with Crowd-
sensed Data. In: Manghi, P., Candela, L., Silvello,
G. (eds) Digital Libraries: Supporting Open Sci-
ence. IRCDL 2019. Communications in Computer
and Information Science, vol 988. Springer, Cham.
https://doi.org/10.1007/978-3-030-11226-4 12
De Vita, F., & Bruneo, D. (2020). Leveraging Stack4Things
for federated learning in intelligent cyber physical sys-
tems. Journal of Sensor and Actuator Networks, 9(4),
59.
Maharshi Dhada, Amit Kumar Jain, and Ajith Kumar Par-
likad. “Empirical Convergence Analysis of Feder-
ated Averaging for Failure Prognosis”. In: IFAC-
PapersOnLine 53.3 (2020). 4th IFAC Workshop on
Advanced Maintenance Engineering, Services and
Technologies - AMEST 2020
Ahmed Gad. Introduction to Federated Learning. Apr.
2020.
Fosca Giannotti, Roberto Trasarti, Kalina Bontcheva, and
Valerio Grossi. 2018. SoBigData: Social Mining &
Big Data Ecosystem. In Companion Proceedings of
The Web Conference 2018 (WWW ’18). International
World Wide Web Conferences Steering Committee,
Republic and Canton of Geneva, CHE, 437–438.
https://doi.org/10.1145/3184558.3186205
General Data Protection Regulation. https://gdpr-info.eu/
Gupta, H., Patel, D., Makade, A., Gupta, K., Vyas, O.
& Puliafito, A. Risk Prediction in the Life Insurance
Industry Using Federated Learning Approach. 2022
IEEE 21st Mediterranean Electrotechnical Confer-
ence (MELECON). pp. 948-953 (2022)
Birjit Gope et al. “Handwritten Digits Identification Using
Mnist Database Via Machine Learning Models”. In:
IOP Conference Series: Materials Science and Engi-
neering 1022.1 (Jan. 2021)
Andrew Hard et al. Federated Learning
for Mobile Keyboard Prediction. 2018.
doi: 10.48550/ARXIV.1811.03604. url:
https://arxiv.org/abs/1811.03604.
Li, T., Sahu, A., Zaheer, M., Sanjabi, M., Tal-
walkar, A. & Smith, V. Federated Optimiza-
tion in Heterogeneous Networks. (arXiv,2018).
https://arxiv.org/abs/1812.06127
Tian Li et al. Federated Optimization in Heterogeneous Net-
works. 2018. arXiv
Tian Li et al. Federated Learning: Challenges, Methods,
and Future Directions. Aug. 2019
Long, G., Shen, T., Tan, Y., Gerrard, L., Clarke, A. & Jiang,
J. Federated Learning for Privacy-Preserving Open In-
novation Future on Digital Health. (2021,8)
McMahan, B., Moore, E., Ramage, D., Hampson, S. &
Arcas, B. Communication-Efficient Learning of Deep
Networks from Decentralized Data. Proceedings Of
The 20th International Conference On Artificial Intel-
ligence And Statistics. 54 pp. 1273-1282 (2017,4,20),
https://proceedings.mlr.press/v54/mcmahan17a.html
Akayuki Nishio and Ryo Yonetani. “Client Selection for
Federated Learning with Heterogeneous Resources
in Mobile Edge”. In: ICC 2019 - 2019 IEEE In-
ternational Conference on Communications (ICC).
IEEE, May 2019. doi:10.1109/icc.2019.8761315.
url:https://doi.org/10.1109%2Ficc.2019.8761315.
Sinno Jialin Pan and Qiang Yang. A Survey on Transfer
Learning”. In: IEEE Transactions on Knowledge and
Data Engineering 22.10 (2010), pp. 1345–1359. doi:
10.1109/TKDE.2009.191.
Scientific LargeScale Infrastructure for Comput-
ing/Communication Experimental Studies Starting
Community. https://slices-sc.eu/.
Scientific LargeScale Infrastructure for Comput-
ing/Communication Experimental Studies.
https://slices-ri.eu/slices-ri-objectives/
SLICE-RI. Scientific LargeScale Infrastructure for
Computing/Communication Experimental Studies.
Url:https://www.slices-ri.eu/slices-ri-objectives/.
Taud, H. & Mas, J. Multilayer Perceptron (MLP). Geomatic
Approaches For Modeling Land Change Scenarios.
pp. 451-455 (2018)
Kang Wei et al. “Federated Learning With Differen-
tial Privacy: Algorithms and Performance Analy-
sis”. In: IEEE Transactions on Information Foren-
sics and Security 15 (2020), pp. 3454–3469. doi:
10.1109/TIFS.2020.2988575.
Empirical Analysis of Federated Learning Algorithms: A Federated Research Infrastructure Use Case
331