Multi-objective Learning in Federation

Federated Learning (FL) is a powerful distributed computing paradigm, developed for scenarios where data originates in distribution and cannot be shared due to privacy, confidentiality, or hardware constraints. Training a machine learning model in such circumstances was previously limited to training a separate model on each data silo, without any knowledge of the data available elsewhere. This approach generally leads to models of lower quality than centralised baselines, trained on a full centralised dataset. By enabling collaborative learning without sharing raw data, Federated Learning promises to unlock the more general insights found in larger datasets for the distributed setting.
Indeed, this strategy has already been widely adopted in the industry, including for deployment on mobile phones and in the finance and health sectors, typically to overcome privacy and confidentiality constraints. However, theoretical challenges remain, often connected to the failure of existing federated algorithms to account for the true complexity of real-world problems. As the capability of machine learning algorithms and hardware grows, so too does the scope for distributed use cases, requiring the adaptation of the federated paradigm to such emerging challenges. Multi-objective modelling is a well-recognised approach to modelling the complexity of the real world with its frequently conflicting requirements. Despite its broad applicability, this direction of research has received very little attention to date. This thesis explores the opportunities and challenges of integrating multi-objective methods with Federated Learning, with a focus on facilitating multi-objective learning in federation. In a first contribution, we provide a comprehensive survey of the literature combining multi-objective and Federated Learning techniques and propose a first systematic taxonomy of the field. We categorise existing works into this taxonomy and identify open areas of research, noting that federated multi-objective learning (FMOL) in particular remains underexplored. Following this insight, we propose a first novel framework and an algorithm, respectively, for two distinct FMOL settings. In the first setting, previously unaddressed in the literature, distributed parties collaborate under the control of a server to find a full spectrum of trade-off solutions. Our proposed framework, MOFL/D, formalises a general approach based on decomposition, a well-established strategy from the field of multi-objective optimisation.
In the second setting, participants assign different pre-defined importance preferences to the objectives of the problem. Each party is interested in finding a single solution that matches its own preferences, leading to the challenge of aligning models with conflicting preferences. We propose an algorithm, FedPref, that finds a personalised model for each participant, modulating collaboration during the learning process based on similarity. Next, we consider how to validate FMOL algorithms appropriately. We argue that the currently predominant benchmarking problems fail to represent the true difficulty of multi-objective learning, lacking inherent conflict between objectives. Consequently, we propose a new class of accessible, flexible, and scalable benchmarking problems, derived from the field of fair machine learning (Fair ML), that are known to contain such conflicts. We demonstrate the instantiation of a range of these Fair ML benchmarks and show their use on state-of-the-art algorithms. In a final chapter, we look towards the future, examining the potential of Federated Learning as a tool in the space domain. We identify a challenging potential use case, envisioning the use of FL algorithms to establish ad-hoc collaboration between satellites performing orbital edge computing. We discuss the state of the art in the field of Federated Learning as relating to this use case, and point out gaps where further research is required. Aside from algorithmic challenges, one crucial gap is a lack of existing standards to facilitate the exchange of machine learning models in spacecraft communications. We consider a potential pathway towards the rapid establishment of such standards.