2019-Survey杨强-安全方面相关工作+横纵向分类-Federated+Machine+Learning-+Concept

12

Federated Machine Learning: Concept and Applications

QIANG YANG, Hong Kong University of Science and Technology, Hong Kong

YANG LI U and TIANJIAN CHEN, Webank, China

YONGXIN TONG, Beihang University, China

Today’s articial intelligence still faces two major challenges. One is that, in most industries, data exists in the

form of isolated islands. The other is the strengthening of data privacy and security. We propose a possible so-

lution to these challenges: secure federated learning. Beyond the federated-learning framework rst proposed

by Google in 2016, we introduce a comprehensive secure federated-learning framework, which includes hori-

Machine learning; Supervised learning;

Additional Key Words and Phrases: Federated learning, GDPR, transfer learning

ACM Reference format:

Qiang Yang, Yang Liu, Tianjian Chen, and Yongxin Tong. 2019. Federated Machine Learning: Concept and

Applications. ACM Trans. Intell. Syst. Technol. 10, 2, Article 12 (January 2019), 19 pages.

https://doi.org/10.1145/3298981

1 INTRODUCTION

2016 is the year when articial intelligence (AI) came of age. With AlphaGo [59] defeating the top

human Go players, we have truly witnessed the huge potential in AI and have began to expect more

complex, cutting-edge AI technology in many applications, including driverless cars, medical care,

will be realized soon in all aspects of our lives. However, real-world situations are somewhat disap-

pointing: with the exception of a few industries, most elds have only limited data or poor-quality

Authors’ addresses: Q. Yang, Hong Kong University of Science and Technology, Clear Water Bay, Kowloon, Hong Kong,

China; email: qyang@cse.ust.hk; Y. Liu and T. Chen, Building #7 A, Keji Shengtaiyuan, No. 1819, Shahe West Road, Nanshan

District, Shenzhen, China; emails: {yangliu, tobychen}@webank.com; Y. Tong (corresponding author), Advanced Innova-

tion Center for Big Data and Brain Computing, Beihang University, No. 37 Xueyuan Road, Haidian District, Beijing, China;

Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee

provided that copies are not made or distributed for prot or commercial advantage and that copies bear this notice and

the full citation on the rst page. Copyrights for third-party components of this work must be honored. For all other uses,

Fig. 1. GDPR: EU regulation on data protection.

data, making the application of AI technology more dicult than we thought. Would it be possible

to fuse the data together in a common site by transporting the data across organizations? In fact, it

is very dicult, if not impossible, in many situations to break the barriers between data sources. In

general, the data required in any AI project involves multiple types. For example, in an AI-driven

product recommendation service, the product seller has information about the product and data

of the user’s purchase but not the data that describes the user’s purchasing ability and payment

habits. In most industries, data exists in the form of isolated islands. Due to industry competition,

privacy security, and complicated administrative procedures, even data integration between dier-

ent departments of the same company faces heavy resistance. It is almost impossible to integrate

example is the General Data Protection Regulation (GDPR) [19]enforcedbytheEuropeanUnion

on May 25, 2018. The GDPR (Figure 1) aims to protect users’ personal privacy and provide data

security. It requires businesses to use clear and plain language for their user agreement and grants

users the “right to be forgotten,” that is, users can have their personal data deleted or withdrawn.

Companies violating the bill will face sti nes. Similar acts of privacy and security are being en-

acted in the United States and China. For example, China’s Cyber Security Law and the General

Principles of the Civil Law, enacted in 2017, require that Internet businesses must not leak or tam-

per with the personal information that they collect and that, when conducting data transactions

with third parties, they need to ensure that proposed contracts follow legal data protection obli-

gations. The establishment of these regulations will clearly help build a more civil society but will

ulations and laws. As well, since users may be unclear about the future uses of the models, the

transactions violate laws such as the GDPR. As a result, we face a dilemma that our data is in the

form of isolated islands, but we are forbidden in many situations to collect, fuse, and use the data

in dierent places for AI processing. How to legally solve the problem of data fragmentation and

isolation is a major challenge for AI researchers and practitioners today.

ACM Transactions on Intelligent Systems and Technology, Vol. 10, No. 2, Article 12. Publication date: January 2019.

In this article, we give an overview of a new approach, known as federated learning, which is a

possible solution for these challenges. We survey existing works on federated learning, and pro-

pose denitions, categorizations, and applications for a comprehensive secure federated-learning

framework. We discuss how the federated-learning framework can be applied to various busi-

devices while preventing data leakage. Recent improvements have been focusing on overcoming

the statistical challenges [60, 77] and improving security [9, 23] in federated learning. There are

also research eorts to make federated learning more personalizable [13, 60]. The above works

all focus on on-device federated learning in which distributed mobile-user interactions are in-

volved and communication cost in massive distribution, unbalanced data distribution, and device

reliability are some of the major factors for optimization. In addition, data is partitioned by user

privacy-preserving machine learning, as reported in [58] because it also considers data privacy in a

decentralized collaborative-learning setting. To extend the concept of federated learning to cover

collaborative-learning scenarios among organizations, we extend the original “federated learn-

system architecture for the federated-learning system.

Dene N data owners {F

}, all of whom wish to train a machine-learning model by con-

solidating their respective data {D

and use D = D

does not expose its data D

to others

, V

. Formally, let δ be a non-negative

real number; if

|< δ, (1)

we say that the federated learning algorithm has δ-accuracy loss.

The denition of data security may dier in dierent scenarios but is required to provide meaningful privacy guarantees.

We demonstrate examples of security denitions in Section 2.3.

ACM Transactions on Intelligent Systems and Technology, Vol. 10, No. 2, Article 12. Publication date: January 2019.