Trust is a prevalent concept in human society, which is particularly associated with situations in which one entity, a truster, needs to rely on the actions of another entity, known as a trustee. Despite the lack of a single accepted definition, trust can be viewed as the subjective probability with which a trustee will act in a certain way, from the point of view of a truster. This notion of trust is not only important in society at large, but it is also becoming increasingly important in the field of computer science. In particular, we are interested in the role that trust plays in service-oriented systems, such as the Semantic Web and the Grid.
A key objective of these systems is to allow computer resources from different geograph- ical locations, or belonging to different organisations, to be used together seamlessly in support of a common goal. The nature of these systems means that resources from organisations that have competing incentives may be used together, and some resource failure should be expected at anytime. As a result, some researchers have suggested that autonomous software agents, which make decisions without human intervention, could
play an important role in managing resources in such environments. However, if this is to be achieved, these agents must be capable of assessing the trustworthiness of their peers.
In response, our aim was to develop trust assessment mechanisms that could be em- ployed by agents in a service-oriented environment. In particular, we identified two major sources of information that these mechanisms should make use of: (1) the direct experiences of a truster with its peers; and (2) third party experience with a trustee, oth- erwise known as reputation. However, the amount of information each of these sources provide may vary depending on the situation; in particular, reputation may not always be reliable, due to the view point and incentives of a truster’s reputation sources. Thus, even though a truster should make use of these sources, it should be able to deal with inaccurate reputation, and give reasonable results regardless of the amount of reputation available.
Before addressing these aims directly, in Chapter 2 we reviewed existing methods in the literature for solving these and similar problems. Here, we saw that previous models differ both in how they represent trust, and in how they reason about it. In terms of representation, some of the prevailing approaches include the application of Dempster- Shafer theory, probability theory, or more improvised methods. Although each of these may have their place, we believe probability is particularly suited in our context for two main reasons. First, assessing the properties of a system based on past behaviour is one of the fundamental questions that probability theory attempts to answer, and which it achieves through a set of well-established techniques with strong theoretical rationales. Second, probability has a natural interpretation in decision theory, which itself is well suited to facilitating decision making by autonomous agents.
There are three main ways by which existing trust models deal with the inherent lack of reliability in reputation. First, a truster may assume that, out of a group of opinions provided about a trustee, only a minority are likely to be inaccurate. The problem with this approach is that, in many situations, this assumption may be inappropriate. For instance, if no agent has any experience of a trustee, any agent that reports having such experience must by lying, and so is not likely to provide useful information. Second, we may try to discourage lying behaviour by designing systems in which it is always in the best interest of an agent to tell the truth. However, this approach may not always be possible, and in any event, cannot deal with inaccuracies due to reasons other than lying. Finally, we may assess the reliability of a particular reputation source by comparing its opinions to subsequent trustee behaviour: the more correlation we observe between such opinions and behaviour, the more reliable the reputation source can be judged to be. However, among existing probabilistic trust models, this approach has not been addressed in a satisfactory manner.
To begin to address these limitations, Chapter 3 set out a framework by which decision theory and Bayesian analysis can be applied to problems involving trust. In particular, it set out a general approach for making decisions based on a truster’s own experiences, and how best to communicate reputation between agents, such that all relevant information is maintained with minimum transmission overhead. Furthermore, to help guide solutions for inaccurate reputation, the chapter categorised the main causes of inaccuracies, along with their effects.
Building on this, Chapter 4 introduced TRAVOS, which instantiates the framework for binary representations of trustee behaviour. TRAVOS includes a mechanism for dealing with inaccurate reputation based on a reputation source’s past performance, and has been applied as part of a larger system for managing resources in a service-oriented environment. Finally, Chapter 5 presented TRAVOS-C, which extends the capabilities of TRAVOS in three ways: (1) by using a continuous representation of trustee behaviour, (2) by including an improved Bayesian mechanism for dealing with inaccurate reputation, and (3) by allowing trustees to be assessed based on the behaviour of similar agents in the system.
6.2
Research Contributions
The main contributions of this thesis stem from the specification of the general frame- work of modelling trust and reputation, and the development of TRAVOS and TRAVOS- C. Together, these show how, by applying standard techniques from statistics and deci- sion theory, an agent can assess the expected benefits of interacting with another agent in a given situation, and so decide which of its peers to interact with, in pursuit of its goals. For example, if an agent has to choose between two providers of a multimedia service, it can use trust to assess how likely each agent is to fulfill its promises, and use trust along with other factors, when making its decisions.
In addition, by applying probability to trust assessment, our methods inherit three key benefits that they share with other probabilistic models of trust. First, by being based on the axioms of probability, these models provide a way of representing beliefs about uncertainty that is consistent and well founded.
Second, by applying well known results, we can derive optimality properties for these mechanisms, under the model assumptions. For instance, decision theory tells us that if an agent has to choose between possible actions, the best choice is always to maximise its expected utility — something which can be directly derived using probability theory. Third, by applying decision theory along with Bayesian analysis, we can meet our ob- jective of making reasonable decisions regardless of the amount of information available. This is because, through Bayesian analysis, we can calculate the marginal distribution
for a trustee’s actions, which accounts for the amount of evidence available in the most appropriate way, given the model assumptions. Then, by applying this in decision the- ory, a truster can make choices that account for both the risks and potential gains of each choice, given the available evidence.
More significantly, we contribute to the state-of-the-art in three main areas: reputation communication, reputation inaccuracy filtering, and trust assessment based on group behaviour. We elaborate on each of these in the subsections that follow.
6.2.1 Communicating Reputation
As part of our general framework (Chapter 3), we specify a set of guidelines for commu- nicating opinions between agents based on direct experience. These act as a benchmark for transmitting reputation between agents, such that if these are met, all relevant in- formation about an agent’s observations are conserved, and this is done with minimum communication overhead. In addition, if these guidelines are adhered to, and an agent’s reputation can be assumed to be accurate, then trust models, based on reputation, can be built to reach conclusions that are consistent and as reliable as conclusions based on direct experience.
These guidelines are fulfilled by a number of models, including TRAVOS, which represent trustee behaviour as a binary event, and are based on the Beta Reputation System (BRS). However, TRAVOS-C is the first trust model to address these guidelines for
continuous representations of trustee behaviour. 6.2.2 Addressing Inaccurate Reputation
In cases where reputation cannot be considered as reliable as direct experience, both TRAVOS and TRAVOS-C implement methods for minimising the impact of inaccu- rate reputation, each of which has its own separate advantages. The method used in TRAVOS works by comparing the past reports of a reputation source about a trustee, with subsequent direct experience with that trustee. Based on this, it calculates the probability that a trustee’s behaviour, on average, lies within a certain margin of error around the reputation source’s best estimate. This is then used as part of a heuristic to mediate each source’s opinions, such that sources whose opinion accuracy lies outside a margin of error will tend to be ignored completely.
This approach has been shown empirically to outperform the only previous method of its kind (Whitby et al., 2004), which operates on the same BRS derived representation of trust. This is especially important when a significant number of a truster’s reputation sources provide inaccurate information, because the method presented by Whitbyet al
TRAVOS has formed the basis of later work, presented by Zhang and Cohen (2006), which extends the technique to include some of the advantages featured by other existing trust models.
Also building on this, TRAVOS-C presents an improved method of reputation filtering that is derived completely from the assumptions of the model, using Bayesian analysis. In this case, interaction outcomes are represented as real numbers that may, for example, be based on quality of service attributes pertaining to a trustee’s performance. Out- comes of interactions between a particular truster and trustee are then assumed to be drawn from a Gaussian distribution with unknown mean and variance. In particular, a truster’s direct observations of a trustee are assumed to be drawn from this distribution, while third party observations are assumed to be drawn from the same distribution, but with added Gaussian noise. Each reputation source is associated with a different noise distribution, which the truster may learn through repeated interactions with both trustees and reputation sources.
This approach has several advantages over both TRAVOS and other filtering methods in the literature. First, by applying Bayesian analysis to the model assumptions, we obtain probability distributions for the model parameters, which are provably correct. As such, the model accounts for all evidence and dependencies between parameters that are correct for the model, and can be used to facilitate choices using decision theory in a manner that is theoretically sound, without the need for heuristics. Of course, this does not imply that the assumptions made are correct for every application, but we have shown empirically that TRAVOS-C is robust against many types of violation in its assumptions, and by making its assumptions explicit, it is clear under what conditions the model operates best.
6.2.3 Assessing Trust based on Group Behaviour
To further improve the assessment of a trustee, TRAVOS-C can judge an agent, based on the behaviour of other similar agents in the system. This method is particularly useful for two reasons. First, if neither a truster or its reputation sources have significant experience with a trustee, then assessment based on group behaviour may still provide a significant improvement over assuming no information at all.
Second, it provides a pragmatic solution to the problem ofwhitewashing, in which agents with a poor reputation attempt to improve their standing, by assuming a new identity. In doing so, an agent effectively wipes out any negative information that its peers have about its behaviour, and so is treated just as any other unknown entity in the system. To deal with this, Zacharia et al. (1999) suggest that newcomers to a system should always be assigned the lowest possible rating. However, this may inhibit good market dynamics, by preventing reliable agents from getting a foothold in the market.
As an alternative approach, Sun et al. (2005) suggest that newcomers should be judged according to the general behaviour of other newcomers to the system. In doing so, we can adapt our assessments according to the proportion of reliable and unreliable agents that enter the system at any one time.
These advantages can also be claimed for existing models that account for group be- haviour, including REGRET (Chapter 2) and Sun et al’s approach. However, these solutions are not directly applicable to probabilistic representations of trust, and require the specification of weights to decide how much impact group behaviour should have. Our approach is significant in that it automatically adapts to the amount of correlation that exists between the behaviour of a group of agents. That is, only if there is evidence that group behaviour is a strong indicator of an individual agent’s behaviour will it have a significant impact on assessment. Conversely, if there a great deal of diversity in the behaviour, then group behaviour will have little or no impact on a truster’s assessment.