Of Strategies and Structures: Motif-Based Fingerprinting Analysis of Online Reputation Networks

Reputation networks are an important building block of distributed systems whenever reliability of nodes is an issue. However, reputation ratings can easily be undercut: colluding nodes can spread good ratings for each other while third parties are hardly able to detect the fraud. There is strong analytical evidence that reputation networks cannot be constructed in a way to guarantee security. Consequently, only statistical approaches are promising. This work pursues a statistical approach inspired by the idea that colluding node's behavior changes the local structure of a reputation network. To measure these structural changes, we extend a graph analysis method originating from molecular biology and combine it with a machine learning approach to analyze fingerprints of node's interactions. We evaluate our method using an adaptive Peer-to-Peer (P2P) streaming system and show that a correct classification of up to 98% is possible.