Towards efficient horizontal federated learning

Preview

PDF WRAP_Theses_Wu_2021.pdf - Unspecified Version - Requires a PDF viewer. Download (9MB) | Preview

Request Changes to record.

Abstract

The advance of Machine Learning (ML) techniques has become the driving force in the development of Artificial Intelligence (AI) applications. However, it also brings about new problems as ML models and algorithms are more data-hungry than ever. First, the sources of data are often geographically distributed and thus gathering all the data for centralised training incurs heavy traffic and prohibitive communication costs. Second and more importantly, moving data out of local devices are now prohibited in many situations due to privacy concerns. As a promising solution, Federated Learning (FL) is a framework proposed by Google for learning from decentralised data without data sharing. Nonetheless, many challenges remain especially regarding how to facilitate efficient FL in different circumstances. The work presented in this thesis is mainly focused on optimising the efficiency of Horizontal Federated Learning, the most commonly used paradigm of FL, from three perspectives: i) redesigning the device– server synchronisation mechanism of FL for failure-tolerant collaborative training over unreliable, heterogeneous end devices (clients), ii) integrating FL with the Mobile Edge Computing (MEC) architecture to utilise the edge layer for robustness, resource efficiency and fast convergence, and iii) optimising the client selection policy in FL through the evaluation of clients’ training value based on the learnt representations of their associated data. Results of extensive experiments are presented to demonstrate the effectiveness of the proposed approaches. In addition to algorithm design and experimental observations, this thesis also discusses the research spectrum of FL and overviews its development in the industry including popular platforms, systems and emerging applications. Based on the work presented, key conclusions and insights are provided with a research outlook for the future study of FL.

Request changes or add full text files to a record

Repository staff actions (login required)

View Item