Multi-agent distributed optimisation in adversarial environments
File(s)
Author(s)
Zhong, Tianyi
Type
Thesis
Abstract
The extensive computational and communication capabilities of Cyber–Physical Systems introduce new challenges in designing distributed algorithms for cooperative and competitive smart devices. A major challenge in distributed systems is ensuring security, particularly in the presence of malicious or self-interested agents. Malicious agents may intentionally inject noise into critical information, such as gradient updates in distributed optimisation processes. Self-interested agents can strategically manipulate the transmitted information to optimise their individual benefits at the expense of global objectives. These behaviours disrupt coordinated actions, thus addressing such issues is crucial for maintaining system efficiency and achieving reliable and resilient optimisation in adversarial environments.
To address these challenges, this thesis is structured around two topics. The first one focuses on designing a resilient filter for convex distributed optimisation algorithms to mitigate the impact of adversarial gradients. The second topic aims at aligning individual and societal interests using incentive mechanisms. We first investigate the cyclical monotonicity property of convex functions, based on which a gradient filter is designed. We discuss that this filter can either be used for detecting adversarial behaviour or for recovering convexity from the attacked data and therefore recovering convergence. We argue that to avoid being detected by the filter, adversarial agents alter their local objective functions and pretend to be regular agents. We then propose an incentive mechanism with an induced game, in which an ε-dominant strategy equilibrium of the system is obtained when all agents truthfully use their local objective functions. We further improve the computational efficiency of the distributed implementation of the mechanism by designing cutting plane-based algorithms. A case study is provided to validate the performance of the proposed filter, mechanism and algorithms.
To address these challenges, this thesis is structured around two topics. The first one focuses on designing a resilient filter for convex distributed optimisation algorithms to mitigate the impact of adversarial gradients. The second topic aims at aligning individual and societal interests using incentive mechanisms. We first investigate the cyclical monotonicity property of convex functions, based on which a gradient filter is designed. We discuss that this filter can either be used for detecting adversarial behaviour or for recovering convexity from the attacked data and therefore recovering convergence. We argue that to avoid being detected by the filter, adversarial agents alter their local objective functions and pretend to be regular agents. We then propose an incentive mechanism with an induced game, in which an ε-dominant strategy equilibrium of the system is obtained when all agents truthfully use their local objective functions. We further improve the computational efficiency of the distributed implementation of the mechanism by designing cutting plane-based algorithms. A case study is provided to validate the performance of the proposed filter, mechanism and algorithms.
Version
Open Access
Date Issued
2025-09-08
Date Awarded
01/02/2026
License URL
Advisor
Angeli, David
Publisher Department
Department of Electrical and Electronic Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)