Section: New Results

Trust

Privacy Preserving Digital Reputation Mechanism

Digital reputation mechanisms have recently emerged as a promising approach to cope with the specificities of large scale and dynamic systems. Similarly to real world reputation, a digital reputation mechanism expresses a collective opinion about a target user based on aggregated feedback about his past behavior. The resulting reputation score is usually a mathematical object, e.g. a number or a percentage. It is used to help entities in deciding whether an interaction with a target user should be considered. Digital reputation mechanisms are thus a powerful tool to incite users to trustworthily behave. Indeed, a user who behaves correctly improves his reputation score, encouraging more users to interact with him. In contrast, misbehaving users have lower reputation scores, which makes it harder for them to interact with other users. To be useful, a reputation mechanism must itself be accurate against adversarial behaviors. Indeed, a user may attack the mechanism to increase his own reputation score or to reduce the reputation of a competitor. A user may also free-ride the mechanism and estimate the reputation of other users without providing his own feedback. From what has been said, it should be clear that reputation is beneficial in order to reduce the potential risk of communicating with almost or completely unknown entities. Unfortunately, the user privacy may easily be jeopardized by reputation mechanisms which is clearly a strong argument to compromise the use of such a mechanism. Indeed, by collecting and aggregating user feedback, or by simply interacting with someone, reputation systems can be easily manipulated in order to deduce user profiles. Thus preserving user privacy while computing robust reputation is a real and important issue that we address in ou work [48] , [52] . Our proposition combines techniques and algorithms coming from both distributed systems and privacy research domains. Specifically, we propose to self-organize agents over a logical structured graph, and to exploit properties of these graphs to anonymously store interactions feedback. By relying on robust reputation scores functions we tolerate ballot stuffing, bad mouthing and repudiation attacks. Finally, we guarantee error bounds on the reputation estimation score.