On the Impact of Witness-Based Collusion in Agent Societies

In ways analogous to humans, autonomous agents require trust and reputation concepts in order to identify communities of agents with which to interact reliably. This paper defines a class of attacks called witness-based collusion attacks designed to exploit trust and reputation models. Empirical results demonstrate that unidimensional trust models are vulnerable to witness-based collusion attacks in ways independent multidimensional trust models are not. This paper analyzes the impact of the proportion of witness-based colluding agents on the society. Furthermore, it demonstrates that here is a need for witness interaction trust to detect colluding agents in addition to the need for direct interaction trust to detect malicious agents. By proposing a set of policies, the paper demonstrates how learning agents can decrease the level of encounter risk in a witness-based collusive society.