Model-based Fraud Detection in Growing Networks
(Presented at the IEEE Conference on Decision and Control in Los Angeles, CA)
People share opinions, exchange information, and trade services on large, interconnected platforms. Because of the size of these platforms, they are common targets for fraudsters who try to deceive randomly selected users. To monitor such behavior, the proposed algorithm evaluates anomalies in the network structure that results from local interactions between users. In particular, the algorithm evaluates the degree of membership to well-defined communities of users and the formation of close-knit groups in their neighborhoods. We identify a set of suspects using a first order approximation of the evolution of the eigenpairs associated to the network; and within the set of suspects, we locate fraudsters based on deviations from the expected local clustering coefficients. Simulations illustrate how incorporating structural properties (their asymptotic behavior) into the design of the algorithm allows us to differentiate between the aggregate dynamics of fraudsters and regular users.