Measuring Attacks on Rating Systems
Intuitively, ratings contain useful information. If we increase the probability that an advisor is an attacker, then we expect the amount of information to go down. In a way, ratings are the dual of privacy, in the sense that the goal is to maximise information, while attackers attempt to minimise the information.
We created an approach that uses probability theory and information theory to measure the amount of information. The simplest model is only suitable for measuring information of individual ratings in isolation. A more advanced model captures the dynamics of an attacker that provides multiple ratings. The current hurdle is to take into account subjective differences between advisors. For several important cases, we can symbolically derive the behaviour of an attacker that minimises the information content of a rating. For other cases, we can numerically approximate the behaviour of an attacker. Moreover, we can identify how robust a system is, by using the minimum amount of information in a rating as a proxy.
Finally, we look ahead, and see how we can effectively use the information that is guaranteed to be present.