Section: New Results

Towards a Trust and Reputation Framework for Social Web Platforms and @-economy

Participants : Thao Nguyen [contact] , Bruno Martin [Unice] , Luigi Liquori, Karl Hanks.

Trust and reputation systems (TRSs) have recently seen as a vital asset for the safety of online interaction environment. They are present in many practical applications, e.g., e-commerce and social web. A lot of more complicated systems in numerous disciplines also have been studied and proposed in academia. They work as a decision support tool for participants in the system, helping them decide whom to trust and how trustworthy the person is in fulfilling a transaction. They are also an effective mechanism to encourage honesty and cooperation among users, resulting in healthy online markets or communities. The basic idea is to let parties rate each other so that new public knowledge can be created from personal experiences. The greatest challenge in designing a TRS is making it robust against malicious attacks. In this paper, we provide readers an overview on the research topic of TRSs, propose a consistent research agenda in studying and designing a robust TRS, and present an implemented reputation computing engine alongside simulation results, which is our preliminary work to acquire the target of a trust and reputation framework for social web applications.

Information concerning the reputation of individuals has always been spread by word-of-mouth and has been used as an enabler of numerous economic and social activities. Especially now, with the development of technology and, in particular, the Internet, reputation information can be broadcast more easily and faster than ever before. Trust and Reputation Systems (TRSs) have gained the attention of many information and computer scientists since the early 2000s. TRSs have a wide range of applications and are domain specific. The multiple areas where they are applied, include social web platforms, e-commerce, peer-to-peer networks, sensor networks, ad-hoc network routing, and so on. Among these, we are most interested in social web platforms. We observe that trust and reputation is used in many online systems, such as online auction and shopping websites, including eBay, where people buy and sell a broad variety of goods and services, and Amazon, which is a world famous online retailer. Online services with TRSs provide a better safety to their users. A good TRS can also create incentives for good behavior and penalize damaging actions. Markets with the support of TRSs will be healthier, with a variety of prices and quality of service. TRSs are very important for an online community, with respect to the safety of participants, robustness of the network against malicious behavior and for fostering a healthy market.

From a functional point of view, a TRS can be split into three components The first component gathers feedback on participants' past behavior from the transactions that they were involved in. This component includes storing feedback from users after each transaction they take part in. The second component computes reputation scores for participants through a Reputation Computing Engine (RCE), based on the gathered information. The third component processes the reputation scores, implementing appropriate reward and punishment policies if needed, and representing reputation scores in a way which gives as much support as possible to users' decision-making. A TRS can be centralized or distributed. In centralized TRSs, there is a central authority responsible for collecting ratings and computing reputation scores for users. Most of the TRSs currently on the Internet are centralized, for example the feedback system on eBay and customer reviews on Amazon. On the other hand, a distributed TRS has no central authority. Each user has to collect ratings and compute reputation scores for other users himself. Almost all proposed TRSs in the literature are distributed.

Some of the main unwanted behaviors of users that might appear in TRSs are: free riding (people are usually not willing to give feedback if they are not given an incentive to do so), untruthful rating (users give incorrect feedback either because of malicious intent or because of unintended and uncontrolled variables), colluding (a group of users coordinate their behavior to inflate each other's reputation scores or bad-mouth other competitors. Colluding motives are only clear in a specific application), whitewashing (a user creates a new identity in the system to replace his old one when the reputation of the old one has gone bad), milking reputation (at first, a participant behaves correctly to get a high reputation and then turns bad to make a profit from their high reputation score). The milking reputation behavior is more harmful to social network services and e-commerce than to the others.

This research aims to build on these studies and systematize the process of designing a TRS in general as depicted in Fig. 12 . First, we characterize the application system into which we want to integrate a TRS, and find and identify new elements of information which substitute for traditional signs of trust and reputation in the physical world. Second, based on the characteristics of the application, we find suitable working mechanisms and processes for each component of the TRS. This step should answer the following questions: “What kind of information do we need to collect and how?", “How should the reputation scores be computed using the collected information?", and “How should they be represented and processed to lead users to a correct decision?". To answer the first question, which corresponds to the information gathering component, we should take advantage of information technology to collect the vast amounts of necessary data. An RCE should meet these criteria: accuracy for long-term performance (distinguishing a newcomer with unknown quality from a low-quality participant who has stayed in the system for a long time), weighting towards recent behavior, smoothness (adding any single rating should not change the score significantly), and robustness against attacks. Third, we study the tentative design obtained after the second step in the presence of selfish behaviors. During the third step, we can repeatedly return to Step 2 whenever appropriate until the system reaches a desired performance. The fourth step will refine the TRS and make it more robust against malicious attacks. If a modification is made, we should return to Step 2 and check all the conditions in steps 2 and 3 before accepting the modification. The paper has been accepted to [22] and an improved software and a full paper are in preparation in 2014.