A Grassroots Approach in P2P Reputation Studies

Reputation is a big topic. Many people from various channels are interested in it, with different approach. In this talk, I\´ll take an approach based on system log data. In particular, I\´ll show how generic reputation models may be mapped onto real P2P systems. This is exemplified by our work around Maze, a large scale non commercial peer-to-peer file sharing system deployed in China by network lab at Peking University. The talk will first address some aspect of user behaviors in a peer-to-peer system like Maze. More specifically, using the logs collected from Maze, three typical kinds of "negative" behaviors are identified, namely free riding, whitewashing, and collusion. The statistics and patterns of them are presented. Moreover, the effectiveness of EigenTrust (a famous P2P reputation algorithm) is tested against real data and we have found that EigenTrust has some difficulties in generating proper trust values for certain peers. Some measures to cure the problems are proposed. Besides, to deeper understand attack behavior, we conducted an attack competition, which attracted more than 600 participants and generated many valuable attack data. As such, a brief account on the analysis of the competition data is also reported.