Designing incentives for peer-to-peer systems

Peer-to-peer systems, networks of egalitarian nodes without a central authority, can achieve massive scalability and fault tolerance through the pooling together of individual resources. Unfortunately, most nodes represent self-interested, or rational, parties that will attempt to maximize their consumption of shared resources while minimizing their own contributions. This constitutes a type of attack that can destabilize the system. The first contribution of this thesis is a proposed taxonomy for these rational attacks and the most common solutions used in contemporary designs to thwart them. One approach is to design the P2P system with incentives for cooperation, so that rational nodes voluntarily behave. We broadly classify these incentives as being either genuine or artificial , with the former describing incentives inherent in peer interactions, and the latter describing a secondary enforcement system. We observe that genuine incentives tend to be more robust to rational manipulations than artificial counterparts. Based on this observation, we also propose two extensions to BitTorrent, a P2P file distribution protocol. While this system is popular, accounting for approximately one-third of current Internet traffic, it has known limitations. Our extensions use genuine incentives to address some of these problems. The first extension improves seeding, an altruistic mode wherein nodes that have completed their download continue to provide upload service. We incentivize seeding by giving long-term identifiers to clients enabling seeding clients to be recognized and rewarded in subsequent downloads. Simulations demonstrate that our method is highly effective in protecting swarms from aggressive clients such as BitTyrant. Finally, we introduce The BitTorrent Anonymity Marketplace , wherein each peer simultaneously joins multiple swarms to disguise their true download intentions. Peers then trade one torrent for another, making the cover traffic valuable as a means of obtaining the real target. Thus, when a neighbor receives a request from a peer for blocks of a torrent, it does not know if the peer is really downloading that torrent, or only using it in trade. Using simulation, we demonstrate that nodes cannot determine peer intent from observed interactions