Strategic and Adaptive Behaviours in Trust Systems

Abstract

Intelligent systems are having significant impact on our daily lives in many ways. These systems can help guide human decisions, act on our behalf and cooperate within mixed-initiative teams. This inter-dependency between humans and AI systems inherently includes a level of risk about the outcomes of actions. To mitigate this risk, the concepts of trust and reputation have received significant attention in multi-agent systems (MASs). Numerous techniques have been proposed to answer the interrelated questions of how to reliably assess the trustworthiness of autonomous systems, how to make robust decisions under uncertainty, and how to establish trust between agents and between agents and humans. Computational models of trust typically focus on evaluating the
trustworthiness of others using direct observations and the opinions of others in order to select partners for delegation or to form and maintain relationships. Significantly less attention, however, has been given to understanding how these systems can reliably operate under budgetary constraints, or their vulnerabilities to external attacks. In this thesis, we propose and evaluate a suite of new decision-making strategies to progressively select trustworthy partners under budgetary constraints. First, we show how this decision-making problem maps to budget-limited multi-armed bandit problems. We then present new decision-making models that incorporate both observations and opinions from others. Finally, we show how these approaches can minimise costs associated with, and the risks involved in interaction with agents with varying and uncertain
reliability.

In order to better understand the performance and reliability of such algorithms, we propose a novel, generic method to automate the process of identifying vulnerabilities in Trust and Reputation systems. We do this by mapping the vulnerability analysis problem to an optimisation problem, and show how efficient sampling methods can be used to search the attack space. We devise an attack model and generate attacks that involve the injection of false evidence to identify vulnerabilities in existing trust models. In this way, we provide an objective means to assess how robust trust and reputation algorithms are to different kinds of attacks and conduct comparative analyses.

More information

Identifiers

ORCID for Taha, Dogan Gunes: orcid.org/0000-0003-2774-7027

ORCID for Timothy Norman: orcid.org/0000-0002-6387-4034

Catalogue record

Export record