2022 IEEE 9th International Conference on Cyber Security and Cloud Computing (CSCloud)/2022 IEEE 8th International

Conference on Edge Computing and Scalable Cloud (EdgeCom)

Quantum Federated Learning:

2022 IEEE 9th International Conference on Cyber Security and Cloud Computing (CSCloud)/2022 IEEE 8th International Conference on Edge Computing and Scalable Cloud (EdgeCom) | 978-1-6654-8066-6/22/$31.00 ©2022 IEEE | DOI: 10.1109/CSCloud-EdgeCom54986.2022.00010

Remarks and Challenges

Harashta Tatimma Larasati1,2 , Muhammad Firdaus3 , and Howon Kim1
1 Pusan National University, Republic of Korea
2 Institut Teknologi Bandung, Indonesia
3 Pukyong National University, Republic of Korea
[email protected], mfi[email protected], [email protected]

Abstract—As the development of quantum computing hard- for practical use, we also discuss the possible approach to
ware is on the rise, its potential application to various research actually incorporate the quantum computation to the practical
areas has been investigated, including to machine learning. federated learning. Note that this paper approaches the issue
Recently, there have been several initiatives to expand the
work to quantum federated learning (QFL). However, challenges from a high-level perspective. Therefore, technical terms are
arise due to the fact that quantum computation poses different omitted, and technical descriptions are outside the scope of
characteristics from classical computation, giving an even more this paper.
challenge for a federated setting. In this paper, we present a high- The rest of this paper is organized as follows. First, we
level overview of the current state of research in QFL. Further- explain the current state of quantum computers in Section II
more, we also describe in brief about quantum computation and
discuss its present limitations in relation to QFL development. before discussing the works on quantum federated learning in
Additionally, possible approaches to deploy QFL are explored. Section III. Subsequently, we discuss the possible approach to
Lastly, remarks and challenges of QFL are also presented. realize quantum federated learning in Section IV. Additionally,
Index Terms—quantum federated learning, quantum comput- several remarks and challenges are also discussed in Section
ing, review, analysis V. Finally, Section VI concludes the paper.
I. I NTRODUCTION II. C URRENT S TATE OF Q UANTUM C OMPUTERS
The notion of Federated Learning (FL) arose as a result A. Quantum Computers and Quantum Algorithms
of the growing developments in personal devices and pri- To the general public, a quantum computer may sometimes
vacy concerns. Federated Learning demonstrates its efficacy be thought of as just the same as a classical computer, but
and privacy preservation via collaborative local training and due to its utilization of the quantum phenomena, it can solve
a shared machine learning model update without exposing problems way faster than classical computers. That is only
individual datasets [1]. Hence, its applications to other fields partially true. A quantum computer does solve a problem faster
have flourished, ranging from 5G [2], 6G [3], urban computing than a classical computer, but only for certain cases in which
[4] to the use in vehicular network and various IoT applications we can formulate the suitable quantum algorithms to solve the
[5]. problem, as well as the corresponding quantum circuit in an
Meanwhile, as the research in the quantum computing see efficient manner.
rapid development in the past years, the use of quantum com- One example of the quantum algorithm proven more ef-
putation (QC) to accelerate machine learning tasks, also known ficient than its classical counterpart is the Shor’s algorithm
as quantum machine learning (QML), have been proposed and [8], which gives an exponential speed-up. The most promi-
gained significant interest. Just very recently, the possibility nent variant Shor’s algorithm reduces the factoring problem
of utilizing QC for federated learning (QFL) has also been to period finding, which can be performed efficiently on a
investigated in a handful of literature, such as in [6], [7]. quantum computer, enabling the cracking of Rivest-Shamir-
Nevertheless, the initiative to employ QC for QFL comes Adleman (RSA) as the widely-used public key cryptosystems,
with an immense challenge, particularly due the fact that the along with the Elliptic Curve Cryptography (ECC)-based
current QC has only been able to support to small tasks due cryptosystem using the other variant of Shor’s algorithm.
to the limitation of the current QC infrastructure, which is
very much still in its early development. In this paper, we B. Current State of Quantum Hardware
provide an overview of the current research in the quantum Regarding the quantum hardware development, despite the
federated learning. Furthermore, we describe the current state significant research efforts and major breakthrough in the
of quantum computing as well as the current issue that past years, current quantum hardware is still in the early
might hinder the development in the actual quantum machine development. It poses challenges (or weaknesses), including
learning, particularly in the context of quantum federated the currently small number of qubits (i.e., the smallest unit
learning. Nevertheless, should there be a time when the of information in a quantum computer), and the difficulty to
quantum computer has finally reached a maturity and ready achieve long coherence time (i.e., how long a quantum system

978-1-6654-8066-6/22/$31.00 ©2022 IEEE 1

DOI 10.1109/CSCloud-EdgeCom54986.2022.00010
Authorized licensed use limited to: Chaitanya Bharathi Institute of Tech - HYDERABAD. Downloaded on January 04,2024 at 03:49:02 UTC from IEEE Xplore. Restrictions apply.
 [16] present QFL for wireless communication using quantum
neural networks (QNN) operations in reducing system com-
plexity. In [17], the FL on-hybrid quantum-classical classifier
is described, where the authors utilized the advantages of
QNN and classical pre-trained convolutional model to compare
between federated and non-federated approach. Based on their
implementation, the result shows a faster distributed training
with slightly smaller trained model accuracy. The work in
[6] focus on implementing QFL framework using Qutip,
a python-based quantum simulator. Moreover, the authors
in [18] present QFL with quantum data to provide a new
quantum dataset with IID and non-IID client’s dataset. Here,
the framework is implemented using TensorFlow Federated
(TFF), TensorFlow Quantum (TFQ), ad Cirq Google sim-
ulator. In [19], the authors comprehensively discuss about
the QFL model based on blind quantum computing archi-
tecture approach and classification problems. The work in
Fig. 1. Illustration of Federated Learning framework, redrawn from [14] [20] proposes a unique Optimized Quantum-based Federated
Learning (OQFL) framework for autonomously adjusting fed-
erated learning hyperparameters in autonomous car contexts
can maintain its quantum state before collapsing to classical utilizing various adversarial approaches. Further, the authors
state, due to interaction with the environments [9]). in [21] discuss about the possibility of quantum computers
Furthermore, the current quantum computers are very prone construction as a promising future technology.
to errors. Hence, two terms are known in the field: logical
qubit (i.e., the ideal qubit without errors), and physical qubits. IV. P OSSIBLE A PPROACH TO R EALIZE QFL
In fact, the number of qubits often mentioned in the quantum A. Motivation
hardware development (e.g., the 127-qubits quantum proces- Even though the actual, practical quantum federated learn-
sors, which is considered the largest superconducting-based ing is very much still in its infancy, as the research in both
quantum computer to date recently achieved by IBM [10]) is quantum machine learning and quantum computing progress,
the physical qubit. Research is still ongoing on how to increase there will come a time when QFL will have a chance to start
the fidelity of the gates (operations in quantum computer, materializing. In this section, we explore the possible method
relatively equivalent to the classical AND or OR gates) to to build QFL with the assumption that a practical quantum
minimize errors, but the errors are still very much prevalent. computer is already available.
Considering the quantum error corrections (QEC) required to For ease of discussion, we start from the discussion of
achieve a fault-tolerant quantum computation (FTQC) [11], to standard, traditional machine learning to finally come to feder-
achieve meaningful computation for large-scale use, e.g., to ated learning to obtain the information of the core difference.
perform Shor’s algorithm for cracking RSA cryptosystem, is Subsequently, we apply the information to the quantum case.
unlikely in the near future.
Fortunately, there are cases where a quantum computer B. Classical: From Traditional Machine Learning to Feder-
with relatively small number of qubits (i.e., 50-100 qubits ated Learning
[12]) and coherence time, as well as the presence of errors, In the conventional machine learning, in the case of client-
can still give meaningful result —and sometimes even shows server architecture, the training process is always done in
advantage than the classical computer —despite the mentioned the server. The server also performs the aggregation of data
challenges. In the time known as the Noisy Intermediate-Scale whereas clients are only as the data provider. However, ma-
Quantum (NISQ) era, quantum devices [12], [13] are predicted chine learning for complex applications necessitates an abun-
to be useful to be explored for various applications, ranging dant quantity of training data, and the input for larger models
from investigation of many-body quantum physics, to machine rises exponentially as the number of parameters increases [1],
learning, known as the quantum machine learning (QML). besides the need for a larger capacity for processing training.
Thus, it has become essential to distribute machine learning
III. W ORKS ON Q UANTUM F EDERATED L EARNING across numerous machines [1]. There comes distributed learn-
To date, there has only been a handful of works in quantum ing, which distributes the training to be not only on the server,
federated learning, which also just only started around the year but also on the devices.
of 2020, such as [15]–[17]. In [15], authors propose Vertical However, this distributed learning approach poses certain
Federated Learning (VFL) with feature extraction approach risks, particularly on the privacy of the users. Federated learn-
to decentralized speech processing to be used in vertical ing is introduced to minimize this risk by keeping the users
federated learning via quantum convolution. The authors of (clients)’ data and perform training within their environment,

Authorized licensed use limited to: Chaitanya Bharathi Institute of Tech - HYDERABAD. Downloaded on January 04,2024 at 03:49:02 UTC from IEEE Xplore. Restrictions apply.
 ෍ realizing a fully-quantum federated learning, a quantum device
will be located at each node, hence all performs certain QML
tasks to accomplish specified objective. In other words, every
Wireless link
node is equipped with a quantum computing capability.
If quantum computing research continue to develop as rapid
as now, in the future there can actually be a fully working and
relatively mature quantum hardware to be used. However, con-
sidering that the types of quantum computers that is prominent
and has progressed rapidly is the superconducting qubit, and
… … the second is the ion-trap based quantum computer, they will
Device
not likely be mobile at first. Hence, the most possible scenario
Local dataset
is to have a quantum-capable device on the server, while
the clients will still run the training classically. Therefore,
the quantum capability or the QML can be used to speed
up training in the server and global aggregation. Later, after
a while when quantum computing capability becomes more
readily available, a fully quantum federated learning can be
accomplished. Figure 3 illustrates the QFL architecture with
the quantum computer in the server, as proposed in [19].
Nevertheless, we can consider the fact that quantum com-
Fig. 2. FL Procedures in Wireless Networks [23] puters can be run remotely, such as what occurs in the
current situation where superconducting qubit-based quantum
hardware of IBM or the ion-trap based quantum computer
and only sends the resulting local model to the server to be
can be accessed from anywhere (of course with the respective
processed and updated as a global model. This process is
authorization). But then, another question will arise, such as
done iteratively until a specified condition is met. Basically,
practicality, or whether the privacy can still be preserved, or
FL aims to facilitate the training model collaboration among
whether the tradeoff of latency can still be tolerated.
participants without conveying their private data; thus, the
Another question that may arise is whether it will be really
private or confidential data is kept and never leaves their
worth it, performance-wise or resource cost-wise, to even
devices [22]. As shown in Figure 1, the FL framework consists
perform it on quantum computer. Nevertheless, the recent
of several steps, i.e., training local model, uploading local
study on QFL mentioned above has shown promising result
model to server, global model aggregation, and delivering a
for a proof-of-concept. Further study and explorations will be
new global model. The FL strives to optimize a global loss
required to finally come to conclusions.
function F(w) through an FL optimization objective that can
be calculated using the empirical risk minimization approach
in Equation 1,
宓宵宬容室宷宨季宇室宷室 宓宵宬容室宷宨季宇室宷室 宓宵宬容室宷宨季宇室宷室
m

min F (w) = pk Fk (w) (1)
w
k=1
…
where w, m, pk , and Fk (w), is notation for model param-
eters, number of devices, number of data points of device k
compared to total number of data points, and loss function of
device k, respectively.
An example of use case and procedure is in the wireless
networks, as shown in Figure 2. As an additional protection
to tackle remaining threat such as membership inference,
research has also leaned towards differential privacy, which
宔宸室宱宷宸宰季宖宨宵容宨宵
can be employed locally in the end device (after local model
training) as a local differential privacy, and on the server
(globally) to form a global differential privacy. Fig. 3. QFL with QC on the server, redrawn from [19]

C. Possible Approach to Realize Federated Learning

V. R EMARKS & C HALLENGES
As can be inferred from the previous section, federated
A. Remarks
learning performs training on two different types of en-
tity/locations: at the server and the clients. In the future, for There are several remarks and critical concerns for im-
plementation of QFL in the current situation, described as

Authorized licensed use limited to: Chaitanya Bharathi Institute of Tech - HYDERABAD. Downloaded on January 04,2024 at 03:49:02 UTC from IEEE Xplore. Restrictions apply.
 follows: an extensive discussion on this aspect. Nevertheless, the
• For the current case, quantum machine learning is expen- authors of [18] have initiated the work on the statistical
sive, and the access to a quantum computer is not always heterogeneity aspect in QFL using quantum data, with
easy (or not always available). both IID and non-IID dataset.
• In the present case, the proof-of-concept shown in the • Approach for beyond supervised learning for real-world
literature uses a very small number of qubits, such as data [24]. As in classical, the task of supervised learning
eight or three, due to the limitation of current system. is less difficult than the unsupervised learning, since it
For running the system for actual use, it is likely that the assumes that all data points are labeled, whereas in the
number of qubits should be much higher. real-world case, most data will likely be unlabeled, or
• Research efforts have shown various advancement for just weakly labeled [24]. In the case of QFL, current
optimizing the resource use, such as the number of qubits. work are also primarily in the supervised case. The usage
To further optimize, research can be directed to the for unsupervised learning will be beneficial to be further
encoding of the classical data, i.e., to find a more efficient investigated.
way to input the data to the quantum circuit so that it can
VI. C ONCLUSIONS
be computed more effectively.
• When the result of the QFL is promising, it can be
In this paper, we have offered a high-level summary of
extended to other case such as in personalized federated the current status of research in QFL. Furthermore, we have
learning, or for accelerating edge intelligence in the provided a brief overview of quantum computation and explore
network. its current limits in regard to QFL development. In addition,
the strategy and prospective circumstances for deploying QFL
B. Challenges were investigated. However, there are several challenges of
Additionally, QFL will also pose challenges which is also current federated learning and quantum computation that need
present in the classical federated learning. As described in to be considered for the future direction.
[24] for the challenge of classical federated learning, there
ACKNOWLEDGMENTS
are several primary challenges which relates to solving the
distributed optimization problem (i.e., Equation 1), inherent This work was supported by Institute for Information &
in a federated learning architecture, as follows: Communications Technology Planning & Evaluation (IITP)
• Expensive communication: Classically, communication is
grant funded by the Korea government (MSIT) (No.2019-
known to be a crucial obstacle for a federated network 0-00033, Study on Quantum Security Evaluation of Cryp-
[25]. For the QFL with a fully-quantum approach, this tography based on Computational Quantum Complexity) and
is even more true. In fact, the different mechanism also supported by Institute of Information & Communications
of quantum mechanics introduce new challenge in the Technology Planning & Evaluation (IITP) grant funded by
network design due to the network’s needs to cater the Korea government (MSIT) (No.2019-0-01343, Regional
quantum phenomenon and to abide by the law of quantum strategic industry convergence security core talent training
mechanics [26]. Distributed quantum computation has business).
been discussed and analyzed by few, such as [27], and R EFERENCES
some of its form has been proposed to be used to run
[1] J. Posner, L. Tseng, M. Aloqaily, and Y. Jararweh, “Fed-
Shor’s algorithm, such as in [28], [29]. Additionally,
erated learning in vehicular networks: Opportunities and
the initiative to establish quantum internet, mentioned
solutions,” IEEE Network, vol. 35, no. 2, pp. 152–159,
as the ultimate phase in the quantum revolution [26],
2021.
hope to achieve a network that can interconnects remote
[2] Y. Liu, J. Peng, J. Kang, A. M. Iliyasu, D. Niyato, and
quantum devices through quantum links in synergy with
A. A. Abd El-Latif, “A secure federated learning frame-
the classical ones. The development on these two may
work for 5g networks,” IEEE Wireless Communications,
be crucial for later adapted for realizing a fully quantum
vol. 27, no. 4, pp. 24–31, 2020.
federated learning.
[3] S. Naser, L. Bariah, S. Muhaidat, et al., “Toward
• System and statistical heterogeneity: In classical system,
federated-learning-enabled visible light communication
system heterogeneity (e.g., hardware variability in pro-
in 6g systems,” IEEE Wireless Communications, vol. 29,
cessor and memory, various types of network connectiv-
no. 1, pp. 48–56, 2022.
ity), including client devices’ capacities and capabilities,
[4] A. Huang, Y. Liu, T. Chen, et al., “Starfl: Hybrid
should be taken into consideration when one wants to
federated learning architecture for smart urban com-
implement a federated learning. Furthermore, statistical
puting,” ACM Transactions on Intelligent Systems and
heterogeneity is as important, due to the difference of
Technology (TIST), vol. 12, no. 4, pp. 1–23, 2021.
ideal setting (e.g., independent and identically distributed
(IID) data) with the real-world setting (i.e., distribution
of clients’ local data is mostly in the form of non-IID
data) [24]. As for the QFL case, there has not been

Authorized licensed use limited to: Chaitanya Bharathi Institute of Tech - HYDERABAD. Downloaded on January 04,2024 at 03:49:02 UTC from IEEE Xplore. Restrictions apply.
 [5] D. N. Gupta, R. Kumar, and S. H. Ansari, “Federated [18] M. Chehimi and W. Saad, “Quantum federated learning
learning for an iot application,” in Federated Learning with quantum data,” in ICASSP 2022-2022 IEEE Inter-
for IoT Applications, S. P. Yadav, B. S. Bhati, D. P. Ma- national Conference on Acoustics, Speech and Signal
hato, and S. Kumar, Eds. Cham: Springer International Processing (ICASSP), IEEE, 2022, pp. 8617–8621.
Publishing, 2022, pp. 53–66, ISBN: 978-3-030-85559- [19] W. Li, S. Lu, and D.-L. Deng, “Quantum federated
8. DOI: 10 . 1007 / 978 - 3 - 030 - 85559 - 8 4. [Online]. learning through blind quantum computing,” Science
Available: https://doi.org/10.1007/978- 3- 030- 85559- China Physics, Mechanics & Astronomy, vol. 64, no. 10,
8 4. pp. 1–8, 2021.
[6] Q. Xia and Q. Li, “Quantumfed: A federated learning [20] W. Yamany, N. Moustafa, and B. Turnbull, “Oqfl: An
framework for collaborative quantum training,” arXiv optimized quantum-based federated learning framework
preprint arXiv:2106.09109, 2021. for defending against adversarial attacks in intelligent
[7] B. Narottama and S. Y. Shin, “Quantum federated transportation systems,” IEEE Transactions on Intelli-
learning for wireless communications,” Journal of the gent Transportation Systems, 2021.
Korean Institute of Communication Sciences, 2020. [21] M. Dyakonov, “When will useful quantum computers
[8] P. W. Shor, “Algorithms for quantum computation: be constructed? not in the foreseeable future, this physi-
Discrete logarithms and factoring,” in Proceedings 35th cist argues. here’s why: The case against: Quantum
annual symposium on foundations of computer science, computing,” Ieee Spectrum, vol. 56, no. 3, pp. 24–29,
Ieee, 1994, pp. 124–134. 2019.
[9] I. L. Chuang, R. Laflamme, P. W. Shor, and W. H. [22] A. Hard, K. Rao, R. Mathews, et al., “Federated learn-
Zurek, “Quantum computers, factoring, and decoher- ing for mobile keyboard prediction,” arXiv preprint
ence,” Science, vol. 270, no. 5242, pp. 1633–1635, arXiv:1811.03604, 2018.
1995. [23] Z. Yang, M. Chen, K.-K. Wong, H. V. Poor, and S. Cui,
[10] P. Ball et al., “First quantum computer to pack 100 “Federated learning for 6g: Applications, challenges,
qubits enters crowded race,” Nature, vol. 599, no. 7886, and opportunities,” Engineering, 2021.
pp. 542–542, 2021. [24] T. Li, A. K. Sahu, A. Talwalkar, and V. Smith, “Fed-
[11] J. M. Gambetta, J. M. Chow, and M. Steffen, “Building erated learning: Challenges, methods, and future direc-
logical qubits in a superconducting quantum computing tions,” IEEE Signal Processing Magazine, vol. 37, no. 3,
system,” npj Quantum Information, vol. 3, no. 1, pp. 1– pp. 50–60, 2020.
7, 2017. [25] K. Bonawitz, H. Eichner, W. Grieskamp, et al., “To-
[12] J. Preskill, “Quantum computing in the nisq era and wards federated learning at scale: System design,” Pro-
beyond,” Quantum, vol. 2, p. 79, 2018. ceedings of Machine Learning and Systems, vol. 1,
[13] K. Oonishi, T. Tanaka, S. Uno, T. Satoh, R. Van pp. 374–388, 2019.
Meter, and N. Kunihiro, “Efficient construction of a [26] A. S. Cacciapuoti, M. Caleffi, F. Tafuri, F. S. Cataliotti,
control modular adder on a carry-lookahead adder us- S. Gherardini, and G. Bianchi, “Quantum internet: Net-
ing relative-phase toffoli gates,” IEEE Transactions on working challenges in distributed quantum computing,”
Quantum Engineering, vol. 3, pp. 1–18, 2021. IEEE Network, vol. 34, no. 1, pp. 137–143, 2019.
[14] J. Xu, B. S. Glicksberg, C. Su, P. Walker, J. Bian, and F. [27] R. Van Meter and S. J. Devitt, “The path to scalable
Wang, “Federated learning for healthcare informatics,” distributed quantum computing,” Computer, vol. 49,
Journal of Healthcare Informatics Research, vol. 5, no. 9, pp. 31–42, 2016.
no. 1, pp. 1–19, 2021. [28] R. D. V. Meter III, “Architecture of a quantum mul-
[15] C.-H. H. Yang, J. Qi, S. Y.-C. Chen, et al., “Decen- ticomputer optimized for shor’s factoring algorithm,”
tralizing feature extraction with quantum convolutional arXiv preprint quant-ph/0607065, 2006.
neural network for automatic speech recognition,” in [29] A. Yimsiriwattana and S. J. Lomonaco Jr, “Distributed
ICASSP 2021-2021 IEEE International Conference on quantum computing: A distributed shor algorithm,”
Acoustics, Speech and Signal Processing (ICASSP), in Quantum Information and Computation II, SPIE,
IEEE, 2021, pp. 6523–6527. vol. 5436, 2004, pp. 360–372.
[16] R. M. Pujahari and A. Tanwar, “Quantum federated
learning for wireless communications,” in Federated
Learning for IoT Applications, S. P. Yadav, B. S. Bhati,
D. P. Mahato, and S. Kumar, Eds. Cham: Springer
International Publishing, 2022, pp. 215–230, ISBN: 978-
3-030-85559-8. DOI: 10.1007/978-3-030-85559-8 14.
[Online]. Available: https://doi.org/10.1007/978-3-030-
85559-8 14.
[17] S. Y.-C. Chen and S. Yoo, “Federated quantum machine
learning,” Entropy, vol. 23, no. 4, p. 460, 2021.

Authorized licensed use limited to: Chaitanya Bharathi Institute of Tech - HYDERABAD. Downloaded on January 04,2024 at 03:49:02 UTC from IEEE Xplore. Restrictions apply.