Attack-Resistance of Computational Trust Models

The World Wide Web encourages widely-distributed, open, decentralised systems that span multiple administrative domains. Recent research has turned to trust management [4] as a framework for decentralising security decisions in such systems. However, whilst traditional security measures such as cryptography and encryption are well-understood (theoretically and empirically), the same cannot be said for computational trust models. This paper describes the attack-resistance of several well-referenced trust models, in a move toward a possible framework and terminology for such analyses. We present a number of open questions, and consider possible future directions in the area. 1 Why Computational Trust Models? The World Wide Web encourages widelydistributed, open systems that span multiple administrative domains. Unfortunately, the characteristics of such systems mean that one cannot rely solely on traditional security measures. These ‘open distributed systems’ have a number of characteristics: • relationships are on a peer-to-peer basis; • many peers will have never previously interacted; • multiple administrative domains; • the lack of any globally trusted third party. The concept of trust management [4] provides a framework for decentralising security decisions, which appears able to provide a different ‘paradigm of security’ in such systems. However, whilst traditional security measures such as cryptography and encryption are well-understood (theoretically and empirically), the same cannot be said for computational trust models. We start by sampling a number of well-referenced trust metrics, then present some terminology for assessing and reasoning about their attack-resistance. 2 A Sampling of Trust Metrics At its heart, a computational trust model contains a trust metric. Reiter and Stubblebine [12] consider the problem of authenticating entities using public-key certification in a large-scale, open, distributed system with no trusted third party to manage the name-key bindings of entities. In this context, a trust model takes as input a set of certificates between keys, a source node and a target node, where the source wishes to determine the name-key binding for the target. A trust metric operates over a certification graph that encodes the trust (certificate) relationships between keys, and returns a trust value which represents how trustworthy the source deems the target name-key binding to be. The problem is this: an attacker wishes to introduce a false name-key binding (to impersonate another entity), known as a forgery. The goal of the trust metric is to resist such attacks by rejecting the forgery. The remainder of this section is devoted to a sampling of trust metrics, concentrating on their attack-resistance. However, trust metrics have a much wider field of application than avoiding forged name-key bindings. We consider trust metrics which operate in the following sense. There is a directed graph G where nodes represent principals and weighted edges represent trust relationships between principles, weighted by a trust value. The metric takes a source and a target principle and determines a trust value between them. As an example, consider recommendation-based trust metrics. The graph is a ‘recommendation graph’ where nodes represent principals and an edge (u, v) with label r means that the current node has a recommendation r from principal u, about principal v. Before investigating trust metrics, we present a brief set of terminology. An attack on a graph G is represented by a new graph G′ which contains at least one new target node, known as the forgery, e.g. a certification graph attack G′ on the set of keys V (known as the victims) allows new or changed edges only from the victims (corresponding to stealing nodes’ secret keys). Figure 1