P2P reputation management: Probabilistic estimation vs. social networksZoran Despotovic*, Karl AbererEcole Polytechnique Fe de rale de Lausanne (EPFL), School of Computer and Communication Sciences, CH-1015 Lausanne, SwitzerlandAvailable online 15 August 2005AbstractManaging trust is a key issue for a wide acceptance of P2P computing, particularly in critical areas such as e-com-merce. Reputation-based trust management has been identified in the literature as a viable solution to the problem. Thecurrent work in the field can be roughly divided into two groups: social networks that rely on aggregating the entire available feedback in the network in hope achieving as much robustness against possible misbehavior as possible and probabilistic models that rely on the well known probabilistic estimation techniques but use only a limited fractionof the available feedback. In this paper we provide first an overview of these techniques and then a comprehensive com- parison of the two classes of approaches. We test their performance against various classes of collusive peer behaviorand analyze their properties with respect to the implementation costs they incur and trust semantics they offer to the decision makers. ? 2005 Elsevier B.V. All rights reserved.Keywords: P2P systems; Trust; Reputation; Social networks; Probabilistic estimation1. IntroductionThe availability of ubiquitous communication through the Internet is driving the migration of commerce and business from direct interactions between people to electronically mediated interac- tions. It is also enabling a transition to peer-to-peer commerce without intermediaries and central institutions, e.g., through P2P networks. However, to have widely accepted e-commerce in P2P net- works one has to eliminate or at least minimize the accompanying risks and threats. They origi- nate primarily in the following requirements: The environment must be open, meaning that the users can join and leave it when they want to. This leaves the users a strong feeling of autonomy and independence and can result in1389-1286/$ - see front matter ? 2005 Elsevier B.V. All rights reserved. doi:10.1016/j.comnet.2005.07.003*Corresponding author. Tel.: +41 21 693 5260; fax: +41 21 693 8115.E-mail address: zoran.despotovicepfl.ch (Z. Despotovic).Computer Networks 50 (2006) 485500www.elsevier.com/locate/comnetvarious misbehavior. The effect is also amplified by many other causes such as the inherent tech- nological decentralization or the lack of per- sonal contact (contextual cues 6). The environment must be decentralized, with- out central points of failure. In particular, it must be free of trusted third parties, that would oversee the transactions and punish or rule out any misbehavior. The environment must be global, implying that well established assurance mechanisms such aslitigation are ineffective due to large transaction costs when crossing jurisdictional borders.One does not have to go this far in order to identify the need for trust management in P2P net- works. Even low level technological issues such as behaving according to the underlying P2P proto- col, e.g., forwarding queries, leave room for much misbehavior 22.Reputation systems 20 offer a viable solution to encouraging trustworthy behavior in P2P net- works. Their key presumptions are that the partic- ipants of an online community engage in repeated interactions and that the information about their past doings is indicative of their future perfor-mance and as such will influence it. Thus, collect- ing, processing, and disseminating the feedback about the participants? past behavior is expected to boost their trustworthiness. Recent empirical studies of eBay?s reputation mechanism (FeedbackForum) confirm this expectation. Namely, 19shows that reputation profiles are predictive of future performance, while 11 and 15 come to the conclusion that Feedback Forum completelyfulfills its promises: the positive feedback of the sellers increases their worth, while the negative one reduces it. There has been a lot of research recently on on- line trust and reputation management. A consider- able fraction of this work targets P2P networksspecifically. Due to the expectation achieving high robustness against a broad range of misbehavior, including various peer collusion patterns, most of the existing approaches aggregate the entire feed- back available in the network in order to assess the trustworthiness of a single node. We term this class of work social networks in the rest of thepaper and describe it in Section 4. Besides social networks, probabilistic estimation methods have been recently proposed as a possible solution. Its key properties are an analytic characterization of the underlying peer behavior in terms of a proba- bility distribution and using well known estimation techniques such as feedback aggregation strate- gies. These are discussed in Section 3. Unlike social networks, probabilistic met