Page 25 - Read Online

Qi et al. Intell Robot 2021;1(1):18-57 I http://dx.doi.org/10.20517/ir.2021.02 Page 20

design.
2. Comprehensive summary for FRL applications. This paper collects a large number of references in the field
of FRL, and provides a comprehensive and detailed investigation of the FRL applications in various areas,
including edge computing, communications, control optimization, attack detection, and some other appli-
cations. For each reference, we discuss the authors’ research ideas and methods, and summarize how the
researchers combine the FRL algorithm with the specific practical problems.
3. Open issues for future research. This paper identifies several open issues for FRL as a guide for further
research. Thescopecoverscommunication, privacyandsecurity, joinandexitmechanismsdesign, learning
convergence and some other issues. We hope that they can broaden the thinking of interested researchers
and provide help for further research on FRL.

The organization of this paper is as follows. To quickly gain a comprehensive understanding of FRL, the paper
starts with FL and RL in Section 2 and Section 3, respectively, and extends the discussion further to FRL in
Section 4. The existing applications of FRL are summarized in Section 5. In addition, a few open issues and
future research directions for FRL are highlighted in Section 6. Finally, the conclusion is given in Section 7.

2. FEDERATED LEARNING
2.1. Federated learning definition and basics
In general, FL is a ML algorithmic framework that allows multiple parties to perform ML under the require-
[7]
ments of privacy protection, data security, and regulations . In FL architecture, model construction includes
two processes: model training and model inference. It is possible to exchange information about the model
between parties during training, but not the data itself, so that data privacy will not be compromised in any
way. An individual party or multiple parties can possess and maintain the trained model. In the process of
model aggregation, more data instances collected from various parties contribute to updating the model. As
the last step, a fair value-distribution mechanism should be used to share the profits obtained by the collabora-
[8]
tive model . The well-designed mechanism enables the federation sustainability. Aiming to build a joint ML
model without sharing local data, FL involves technologies from different research fields such as distributed
[9]
systems, information communication, ML and cryptography . FL has the following characteristics as a result
of these techniques, i.e.,

• Distribution. There are two or more parties that hope to jointly build a model to tackle similar tasks. Each
party holds independent data and would like to use it for model training.
• Data protection. The data held by each party does not need to be sent to the other during the training of
the model. The learned profits or experiences are conveyed through model parameters that do not involve
privacy.
• Secure communication. The model is able to be transmitted between parties with the support of an encryp-
tion scheme. The original data cannot be inferred even if it is eavesdropped during transmission.
• Generality. It is possible to apply FL to different data structures and institutions without regard to domains
or algorithms.
• Guaranteed performance. The performance of the resulting model is very close to that of the ideal model
established with all data transferred to one centralized party.
• Status equality. To ensure the fairness of cooperation, all participating parties are on an equal footing. The
shared model can be used by each party to improve its local models when needed.

A formal definition of FL is presented as follows. Consider that there are parties {F } =1 interested in es-
tablishing and training a cooperative ML model. Each party has their respective datasets D . Traditional ML

approaches consist of collecting all data {D } =1 together to form a centralized dataset R at one data server.
The expected model M is trained by using the dataset R. On the other hand, FL is a reform of ML process
in which the participants F with data D jointly train a target model M without aggregating their data.
Respective data D is stored on the owner F and not exposed to others. In addition, the performance mea-