Resource allocation for federated learning enabled edge intelligence

• Main Author: Lim, Bryan Wei Yang
• Other Authors: Miao Chun Yan
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Computer science and engineering; Engineering::General
• Online Access: https://hdl.handle.net/10356/156418
• Institution: Nanyang Technological University

The confluence of Edge Computing and Artificial Intelligence (AI) has driven the rise of Edge Intelligence, which leverages the storage, communication, and computation capabilities of end devices and edge servers to empower AI implementation at scale closer to where data is generated. An enabling technology of Edge Intelligence is the privacy-preserving machine learning paradigm known as Federated Learning (FL). In FL, end users carry out model training locally before transmitting the model parameters or gradient updates, rather than the raw data, to a model owner for aggregation. Amid the increasingly stringent privacy regulations, FL has enabled the development of applications that have to be built using sensitive user data and will continue to revolutionize service delivery in the Finance, Internet of Things (IoT), healthcare, and transport industries, among others. However, the implementation of FL is envisioned to involve thousands of heterogeneous distributed end devices that differ in terms of communication and computation resources, as well as the levels of willingness to participate in the collaborative model training process. The potential node failures, device dropouts, and stragglers effect are key bottlenecks that impede the effective, sustainable, and scalable implementation of FL. In this thesis, I will first present a tutorial and survey on FL and highlight its role in enabling Edge Intelligence. This tutorial and survey provide readers with a comprehensive introduction to the forefront challenges and state-of-the-art approaches towards implementing FL at the edge. In consideration of resource heterogeneity, I then provide multifaceted solutions towards improving the efficiency of resource allocation for implementing FL at scale amid information asymmetry. In the first study, I propose a vanilla contract-theoretic optimization approach towards balancing the tradeoffs of information freshness and service latency in a federated crowdsensing scenario. In the second study, I devise a multi-dimensional contract-matching approach for optimized resource allocation amid multiple sources of heterogeneities. The first two studies leverage the self-revealing properties of contract theory to solve the resource allocation problem in the information asymmetric edge network. Through performance evaluation, it is shown that even when the worker types are unknown, the resource allocation in the edge network is optimized. In the third study, in face of bounded rationality and dynamic decisions of the workers, I propose a two-level evolutionary game theoretic and auction approach to allocate and price resources to facilitate efficient edge intelligence. The performance evaluation shows the convergence and uniqueness of solutions, as well as the profit maximization aspect of the proposed solution. The studies presented in the thesis are formulated via the interdisciplinary interplay of concepts derived from network economics, optimization, game theory, and machine learning. Finally, I outline promising research directions for future works.