Is Guilt by Association a Bad Thing?

In this paper we study a classification model that mimics guilt by association. We consider a population consisting of known adversaries, covert adversaries, and benign individuals. In our model we associate a suspicion score with each individual. The scores for the covert and benign populations are initially set to 0, while the known adversaries have a fixed score of 1. The scores change dynamically as individuals meet each other. An individual will be classified as an adversary if his score exceeds a certain threshold. We demonstrate analytically and empirically that a method for assigning suspicion scores, and hence the classification procedure itself, can be very sensitive to its parameters. This suggests that guilt-by-association models, and classification procedures based on them, should be treated very carefully and analyzed for robustness to their parameters.