Using Information Theory to Improve the Robustness of Trust Systems

Unfair rating attacks to trust systems can affect the accuracy of trust evaluation when trust ratings (recommendations) about trustee agents are sought by truster agents from others (advisor agents). A robust trust system should remain accurate, even under the worst-case attacks which yield the least useful recommendations. In this work, we base on information theory to quantify the utility of recommendations. We analyse models where the advisors have the worst-case behaviour. With these models, we formally prove that if the fraction of dishonest advisors exceeds a certain threshold, recommendations become completely useless (in the worst case). Our evaluations on several popular trust models show that they cannot provide accurate trust evaluation under the worst-case as well as many other types of unfair rating attacks. Our way of explicitly modelling dishonest advisors induces a method of computing trust accurately, which can serve to improve the robustness of the trust models.