Impact of Feedback on Trust in P2P Networks

Transcription

1 82 JOURNAL OF NETWORKS, VOL. 7, NO. 8, AUGUST 22 Impact of Feedback on Trust in P2P Networks Zhihua Huang ),2), Songnian Lu ), Aixin Zhang 3), Jie Gu ) ) Department of Electronic Engineering, Shanghai Jiao Tong University, Shanghai, China 2) School of Information Science and Engineering, Xinjiang University, Urumqi, China 3) School of Information Security Engineering, Shanghai Jiao Tong University, Shanghai, China zhihua.huang.cn@gmail.com, {snlu,axzhang,ee gujie}@sjtu.edu.cn Abstract Feedback plays an important role in P2P trust systems. Existing trust systems usually assume that most of normal peers have idealized feedback behaviors, that is, peers can provide honest feedback after download immediately. However, studies show that feedback sparsity and delay widely exist in a real P2P network. Unfortunately, their effect on trust systems has received little attention in previous researches. In this paper, we propose a feedback model considering feedback ratio and feedback delay to analyze the effect. We find that most trust systems have poor reliability when the number of feedbacks is small in a decentralized P2P network. We propose an implicit feedback mechanism based on the retention time of files to remove the impact of user feedback on trust systems. Simulation results show that trust systems using implicit feedback can not only effectively isolate normal peers from malicious peers but also provide differential services for normal peers with different behaviors. Index Terms P2P network, trust, feedback sparsity, feedback delay, implicit feedback, performance evaluation I. INTRODUCTION In the open and self-organizing peer-to-peer (P2P) environment, peers act as clients and request services from other peers as well as servers and provide services to other peers. As there is no central authority for identity and content authentication, selfish peers may act in their interests not offering their resources for sharing [], and malicious peers may easily inject unauthentic or malicious contents into the networks [2]. These misbehaviors greatly reduce the overall performance of P2P system. In recent years, a number of trust systems such as EigenTrust [3], PowerTrust [4], PeerTrust [5], have been proposed to protect P2P networks against the increasing misbehaviors of peers. In these systems, each peer is supposed to provide feedback to system after transaction. The effectiveness of trust mechanisms depends on sufficient honest feedbacks provided by peers [6]. However, in a real P2P network, not all pees provide feedback as expected. Liang et al. [2] find that even if KaZaA provides incentive mechanism for feedback, the number of users providing feedback is surprisingly little. As the cost of time and effort is needed for feedback, many users would not like to be involved. Lee et al. [7] find that some users do even not timely check the quality of the downloaded content. Unfortunately, most of trust mechanisms focus on This work was supported by the National 973 Program of China under Grant No. 2CB7343. studying trust algorithms based on the idealized feedback assumption. The characteristic of user feedback in a real P2P network has often been ignored. We raise two issues. Are the existing trust systems still effective when the idealized assumptions do not hold? How to evaluate the quality of transaction if most of users do not provide feedback? For the first issue, we model the real feedback behaviors of normal peers in P2P trust systems and analyze the performances of several typical global and personalized trust systems based on the feedback model. We find feedback sparsity and delay may significantly reduce the effectiveness and reliability of almost all trust systems. For the second issue, we propose an implicit feedback mechanism based on the retention time of the downloaded file, removing the effect of user feedback on trust systems. Under this way, the rater rates the other peers based on the duration that he remains the files downloaded from them in the shared folder. Using implicit feedback as the input data of trust systems, we reassess the performance of the trust systems above in the same environment. The results show that the implicit feedback can make the trust systems play their roles better than explicit feedback. Furthermore, implicit feedback mechanism can encourage peers to behave as expected by system by providing differential services for various peers. Our contributions include three folds: ) We model feedback behavior in a real P2P network and analyze the performance of several typical trust systems based on the feedback model. 2) We propose an implicit feedback mechanism based on the retention time of the downloaded file, removing the effect of explicit feedback on trust systems. 3) We compare the performance of trust systems based on implicit feedback and explicit feedback in the decentralized P2P network by simulation and get some interesting conclusions. The remaining parts of the paper are organized as follows. In section 2, we review several typical trust algorithms in the decentralized P2P network and propose a feedback model. In section 3, we propose an implicit feedback mechanism for P2P file sharing systems. In section 4, we analyze the impact of feedback sparsity and delay on explicit feedback-based trust systems in a uniform simulation environment. Furthermore, we also compare the performances of trust systems using explicit doi:34/jnw

2 JOURNAL OF NETWORKS, VOL. 7, NO. 8, AUGUST feedback and implicit feedback. Finally, we conclude the paper in Section 5. II. FEEDBACK-BASED TRUST MECHANISMS In the decentralized P2P networks, since there is no central entity to manage trust, peers have to cooperate to collect feedbacks, compute and store trust values. Generally, the trust value is calculated as a function of integrating multiple factors such as feedback, file size, bandwidth, and the time of interaction, etc. Since our emphasis is to analyze the impact of feedback on trust systems, for simplicity, the other factors are not taken into account in this paper. That is to say, feedback is the unique input data of the trust systems. Assuming that a peer always provides a binary feedback whenever an interaction occurs, let sat ij and usat ij represent the number of the positive and negative feedbacks about peer j given by peer i respectively, the direct trust DT ij placed in peer j by peer i can be computed by summing the number of positive feedbacks and negative feedbacks separately, and to keep a total score as the positive score minus the negative score [3]. To ensure the trust value is between - and, the direct trust can be normalized as DT ij (t) = sat ij(t) usat ij (t) sat ij (t) + usat ij (t) In a large scale P2P network with millions of peers, each peer has limited interactions with other peers, so each peer can only know the trust values of few peers based on individual experiences. Recommendation and reputation algorithms are usually used to infer trusts in stranger peers. Many recommendation and reputation systems have been proposed [8]. Because of space limitation we only analyze several representative trust systems. A. Several decentralized trust systems Depending on the information source of trust, we can classify trust systems into global trust and personalized trust. ) Global trust systems In the global trust systems, each peer has a global reputation which is computed by the local trust values assigned to it by other peers, weighted by the global reputations of the assigning peers. In all global reputation systems proposed, EigenTrust, PowerTrust, and PeerTrust are the three most famous algorithms. EigenTrust [3] presents a global trust method to minimize the impact of malicious peers on the performance of a P2P system. It is supposed that all peers provide feedback after transaction and the peer who provides authentic content always provides honest feedback. Eigentrust can effectively identify malicious peers and isolates them from the network when all users provide feedbacks. Some pretrusted peers are needed for ensuring the reliability of trust. () PowerTrust [4] analyzes the feedback properties of ebay reputation system and points out that the number of feedbacks received by peers follows the powerlaw distribution. They prove that the feedbacks of P2P reputation systems have the same properties. The peers with a large number of feedbacks are selected as power peers to replace the pre-trusted peers in Eigentrust. The operational efficiency comes mainly from the reliable power peers. PeerTrust [5] computes a peer s reputation based on the number of transactions, the size of each transaction, the number of feedbacks, and the credibility of feedback. The credibility of feedback may be measured by two ways: ) a function of the trust value of the feedback provider; 2) the personalized similarity of the feedback provider and the trust requester in rating other peers. We only analyze the second way in this paper. To encourage peers to provide feedback, PeerTrust takes the feedback ratio as one of the recommendation factors. The trust systems above build trust based on the public opinion. In the decentralized P2P networks, computing peers global reputation is complicated and expensive and involves a lot of communication and calculation. Some personalized trust mechanisms have been proposed. 2) Personalized trust systems In the personalized trust systems, the trust placed by a peer in the other peers is built on the direct experience or the opinions from a group of other peers selected by the peer. Liang et al. [9] model the decentralized trust by the following equation: T = β DT + ( β) RT (2) where T denotes the personalized trust value, DT denotes the trust based on self-experience information, RT denotes the recommendation trust from other peers. Generally, the direct experience plays a more important role than the indirect experience in trust systems. We set β =.7. Several different recommendation algorithms are compared in the same context and platform. They find that the simple average aggregation and beta reputation algorithms perform better than the complex algorithms when there are a considerable number of bad raters in the system. Therefore, we only select the following three algorithms for feedback analysis. We simply describe these algorithms, and the detailed introduction may refer to the original literatures. Liang et al. [] propose an average aggregation algorithm (AV G) to compute the recommendation. Let DT kj denote peer k s rating towards peer j, s denote the number of peers who rated peer j, RT ij denote the value of aggregation ratings to peer j in peer i, RT ij can be expressed as follow: s k= RT ij = DT kj (3) s Wang et al. [] propose a half weighted algorithm (HALF ) to compute recommendation. Let A i is the set of acquaintance peers and S i is the set of stranger peers for

3 84 JOURNAL OF NETWORKS, VOL. 7, NO. 8, AUGUST 22 peer i, w a and w s denote the weight of acquaintance recommendation and stranger recommendation respectively, W ik is the weight about the rating from the acquaintance peer k, RT ij can be expressed as RT ij = w a k A i W ik DT kj k A i W ik + w s W ik is updated by the equation: x S i DT xj S i (4) W ik = α W o ik + ( α) e α (5) where W ik denotes the new trust value placed by peer i in peer k, Wik o denotes the old trust value, α is the learning rate, e α is the new evidence value. Here we simplify W ik as DT ik, that is, the rating of an acquaintance of peer i is weighted by peer i s direct trust in it. Intuitively, the opinion of the acquaintance peers are more credible than that of the stranger peers. Josang et al. [2] use beta probability density function to combine rating and derive reputation (BET A). Let pr and nr denote the number of positive and negative ratings on peer j respectively, the recommendation trust in peer j can be expressed as RT ij = pr + pr + nr + All trust mechanisms above take feedbacks as the input data of the trust systems. When all normal users immediately provide honest feedback after transaction, the effectiveness of trust systems has been verified. However, Liang [2] and Lee [7] point out the idealized feedback assumption is not true in a real P2P network. It is necessary to analyze the performances of trust mechanisms under the real feedback behaviors. Next, we model the real feedback behavior through feedback ratio and feedback delay. B. Modeling feedback behaviors We model a P2P file sharing network by a tuple G = (P, E, C, B) where P is the set of peers in P2P network, P = {p,..., p n }, E is the set of neighbor relationships between peers, E = {e ij i j, i, j P }, C is the set of resources provided by all peers, and B is the set of user behaviors. We consider two kinds of user behaviors: sharing and feedback, described by B = {SHARE, F B}. SHARE represents the contributions of all peers to P2P service system while F B represents the contributions of all peers to the rating system. Since P2P network is highly dynamic, user behavior changes over time t. A peer s sharing behavior can be described by the set of files shared or transacted by it. Let C(j, t) denote the set of files shared or transacted by peer j up to time t, C(j, t) C. Each file in C(j, t) has three properties: UV jf (t) represents the total volume of file f uploaded by peer j up to time t; SD jf (t) represents the duration that peer j shares file f up to time t; P R f represents the identification of the peer who provided file f for peer j. Let SHARE(j, t) denote the (6) sharing history of peer j up to time t, SHARE(j, t) = {f, UV jf (t), SD jf (t), P R f f C(j, t)}. The sharing histories of all peers can be described by SHARE(t) = {SHARE(j, t) j P }. Let F B(j, t) represent the feedback history of a peer j up to time t, P (j, t) denote the set of peers who interacted with peer j up to time t, DT ji (t) denote peer j s rating on peer i up to time t, F B(j, t) = {DT ji (t) i P (j, t)}. The feedback histories of all peers can be described by F B(t) = {F B(j, t) j P }. Let fr denote feedback ratio that is the ratio of feedbacks to downloads, F jf (t, fd, fr) denote the feedback result that peer j provides with probability fr at time t about file f downloaded at time t fd, F jf (t, fd, fr) {,, }, peer j s rating on peer i at time t can be expressed as t DT ji (t) = DT ji (t ) + F jf (t, fd, fr) fd= f C(j,t) P R f =i where fd denotes the delay time from completing download to checking the downloaded file. When fr =, peer j provides no feedback, F jf (t, fd, fr) =. According to [7], the delay time shows a bimodal form, that is, one part of users immediately check the file after download while another part of users check the file until finishing session. It is reasonable to suppose that a peer provides feedback only when checking the quality of files. Therefore, the delay time of providing feedback follows the same distribution as the checking delay time. In section IV, we analyze the performance of trust systems with various feedback ratios and feedback delays. III. AN IMPLICIT FEEDBACK MECHANISM In a decentralized P2P network, the trust systems based on explicit feedback are difficult to implement if the majority of users are unwilling to provide feedback information for the items they download. Some implicit rating methods are proposed to infer the credibility of an item through some objective factors. Feng et al. [3] use a lifetime and popularity based ranking approach to generate the rating on a file. Fotiou et al. [4] take the fact that a peer shares an information version as a positive vote on a file. Ormandi et al. [5] infer implicit ratings on files by observing whether the rater will seed the file. If a user deletes a file after download, his rating on the file is low, or if a user is seeding the file for a long time, his rating is high. However, how to implicitly rate a peer still remains an open question. When a peer completes download, its behavior expected by system is to immediately check the quality of the downloaded file, delete it if the file is polluted or continue to share it if the file is authentic. Assuming that all peers behave as expected, the evaluation of a peer on the downloaded file may be inferred from the retention time of the file in shared folder after download. Similarly, the evaluation of a peer on the other peers may be inferred from the average retention time of the files downloaded (7)

4 JOURNAL OF NETWORKS, VOL. 7, NO. 8, AUGUST from them. Based on this idea, we propose an implicit feedback mechanism. Let F S j (t) denote the set of files downloaded by peer j up to time t, RS j (t) denote the set of the files which are evaluated up to time t, RS j (t) F S j (t). Let IT jf (t) denote the interval from the moment when peer j completes downloading file f to the current time t, SD jf (t) denote the duration that peer j remains file f in the shared folder during IT jf (t), the implicit rating on peer i placed by peer j can be inferred by the following rules: ) If IT jf (t) > T H, put file f in RS j (t), or else peer j s rating on file f is not considered. Here T H is the time threshold of the valid rating. It may be set by each user or be a experience value. We assume that T H equals to three times slots in this paper. 2) If RS j (t), let RSP ji (t) denote the set of files provided by peer i in RS j (t), ASD ji (t) denote the average duration that peer j remains files provided by peer i in RS j (t), then f RSP ASD ji (t) = SD ji(t) jf (t) (8) RSP ji (t) 3) Let P SET j (t) denote the set of peers who provided files in RS j (t) for peer j up to time t, peer j s implicit rating on peer i in P SET j (t) can be computed by IDT ji (t) = ASD ji (t) k P SET ASD j(t) jk(t) Thus at time t, the system can get the feedback information of all peers who interact with other peers at time t T H. Under the implicit feedback scheme, a peer will be punished by the declining quality of service in the future if it does not behave as expected. For example, if peer j remains the polluted files downloaded from peer i in the shared folder for a long time, from the perspective of system, peer j s rating on peer i may be high. This will improve the probability that peer j interacts with peer i in the future. Therefore, the implicit feedback mechanism has both reward and punishment function. It can encourage peers to actively check the quality of the downloaded file and filter out invalid files as well as forward the authentic files. Implicit feedback can be used in all existing trust systems. In next section, we analyze the performance of various trust systems using implicit feedback. A. Simulation settings IV. EVALUATION In order to analyze the impact of feedback on the trust systems, we simulate a decentralized P2P trust network with n peers, in which various trust mechanisms are used to improve the download success rate of normal peers. We compare the performance of trust mechanisms in the networks with the idealized feedback behavior and with feedback sparsity and delay. Simulation environment consists of five parts as follows. (9) Network model. We simulate a P2P network with small world topology. Assuming that each peer may connect d max peers at most and d min peers at least. Initially, peer j is randomly assigned d j (d min d j d max ) peers as neighbors. With the evolution of trust system, a peer may select the d max most trustworthiness peers as neighbors. Content model. There are m available files in the network, including z topics with valid and invalid versions each. The popularity of topics and versions of each topic obeys Zipf distribution among the n peers [6]. Here we define the popularity of a topic as the percent of users querying the topic in all users. If a peer has a version file of a topic, we say it has one copy of the version file. For simplicity, we assume that a peer only remains one version file of a topic in the shared folder. Intuitively, the more popular files are more vulnerable to attack. To close to a real network, we assume that the number of the polluted version files of a topic proportionate to the popularity of the topic. A version file f can uniquely be identified by an identification F ID. Peer model. The population of network consists of normal and malicious peers. Normal peers refer to those who have no intention to attack P2P systems. They share valid version files. Each normal peer is interested in a subset of all topics. A peer only shares and search for the files about the interested topics. The popularity of interests also obeys Zipf distributions. We consider two kinds of feedback behaviors of normal users: ) All normal peers behave as expected by system, that is, immediately check the content after download, delete polluted files, continue to share the authentic files, and provide honest feedback. This situation is denoted by fr =, fd =. 2) All normal peers delay fd time slots to check the content after download and provide feedbacks with the probability of f r =. The variable f d obeys the bimodal distribution over [, L] where L is the duration of sharing the file after download [7]. The value of L follows the exponential distribution with parameter is 3. This situation is simply described by fr =, fd = 3. Each malicious peer has an attack target that is a subset of all topics. Malicious peers share the polluted version files in the attack target set.to attack the target, malicious peers always provide false feedbacks without any delay. Simulation execution. The simulation of a P2P network proceeds in time slots. Each time slot is subdivided into a number of query cycles. In each query cycle, a peer issues the query to all his neighbors for a file in his interested subset, and decides whether to respond to queries passing by. The query is propagated by broadcast within time to live (T T L). Upon issuing a query, the peer waits for the responses, and selects a file provided by the most trusted peer to download. After download, the peer may delay a certain time to check the content and provide feedback with a certain probability. Each version file is downloaded at most once by a peer. Upon the conclusion of each query cycle, the system updates the trust value

5 86 JOURNAL OF NETWORKS, VOL. 7, NO. 8, AUGUST 22 TABLE I. SIMULATION SETTINGS Parameters Settings n 2 m 2 z 2 T T L 3 d max d min 3 s the number of query cycles for each slot of each peer. Statistics are collected at each peer. Each experiment is run five times and the results of all runs are averaged. The simulation parameters are shown in Table I. Metrics. We analyze the performance of the trust systems by the following metrics. ) Effective download ratio. It refers to the ratio of good downloads to the total searches for normal peers. 2) Pollution ratio. It refers to the ratio of the polluted copies to the total copies in the system. B. Simulation results ) The impact of feedback on trust systems. We first compare the performance of three typical global trust systems in the networks with the idealized feedback (fr =, fd = ) and with feedback sparsity and delay (fr =, fd = 3). Fig. and Fig. 2 show the evolution of effective download ratio and pollution ratio. When the number of normal feedbacks is sparse, the effective download ratio of three global trust systems decreases while the pollution ratio increases. By comparison, EigenTrust is affected less than PowerTrust and PeerTrust. This is because EigenTrust takes some reliable normal peers as the pre-trusted peers who are selected by normal peers for transactions when without feedback. In PowerTrust, the trust of each peer is aggregated from local trust scores weighted by the global reputations of all raters. When there are a large number of malicious peers in system, they may promote reputation for each other and play more important roles than normal peers in trust evaluation. PeerTrust are the most affected by feedback sparsity and delay. Sparse feedbacks result in the shortage of ratings, which means that the similarity of most peers is unknown. Feedback ratio becomes the important factor of recommendation. Since malicious peers always provide feedback, malicious recommendation will be adopted more. Fig. 3 and Fig. 4 show the performance of personalized trust systems. We can see that feedback sparsity and delay reduce effective download ratio and increase pollution ratio in three personalized algorithms. The AVG and HALF algorithms are affected more than BETA algorithm. The AVG algorithm considers that the weights of all raters are equal. When the number of malicious raters is more than that of normal raters, normal peers will trust in malicious Effective download ratio.9.7. Eigentrust, fr=,fd= Eigentrust, fr=,fd=3 Powertrust, fr=,fd= Powertrust, fr=,fd=3 Peertrust, fr=,fd= Peertrust, fr=,fd= Figure. Comparing the effective download ratio of global trust systems with idealized feedback and with feedback sparsity and delay. Pollution ratio 5 5 Eigentrust, fr=,fd= Eigentrust, fr=,fd=3 5 Powertrust, fr=,fd= Powertrust, fr=,fd=3 Peertrust, fr=,fd= Peertrust, fr=,fd= Figure 2. Comparing the pollution ratio of global trust systems with idealized feedback and with feedback sparsity and delay. Effective download ratio.9.7. AVG, fr=,fd= AVG, fr=,fd=3 HALF, fr=,fd= HALF, fr=,fd=3 BETA, fr=,fd= BETA, fr=,fd= Figure 3. Comparing the effective download ratio of personalized trust systems with idealized feedback and with feedback sparsity and delay. peers. The HALF algorithm gives the acquaintance peer higher weight (here is.7). When few normal peers provide feedbacks, the number of acquaintance peers is very small. The performance of HALF is dependent on the stranger recommendation more. In summary, feedback sparsity and delay can significantly reduce the performance of trust systems. It is

6 JOURNAL OF NETWORKS, VOL. 7, NO. 8, AUGUST Pollution ratio 5 5 AVG, fr=,fd= AVG, fr=,fd=3 HALF, fr=,fd= HALF, fr=,fd=3 BETA, fr=,fd= BETA, fr=,fd= Pollution ratio 5 AVG, fr=,fd= AVG, fr=,fd=3 HALF, fr=,fd= HALF, fr=,fd=3 Peertrust, fr=,fd= Peertrust, fr=,fd= Figure 4. Comparing the pollution ratio of personalized trust systems with idealized feedback and with feedback sparsity and delay. Figure 6. Comparing the pollution ratio of three trust systems with implicit feedback. AVG HALF.9 Effective download ratio.7 AVG, fr=,fd= AVG, fr=,fd=3 HALF, fr=,fd= HALF, fr=,fd=3 Peertrust, fr=,fd= Peertrust, fr=,fd= Figure 5. Comparing the effective download ratio of three trust systems with implicit feedback. necessary to study how to get sufficient input data for trust systems. 2) Using the implicit feedback in trust systems To validate the trust systems with implicit feedback, we compare the performances of the most affected-byfeedback AVG, HALF, and PeerTrust algorithms under implicit feedback mechanism. Since implicit feedback does not need user participation, the feedback ratio does not affect the performance of trust systems. Fig. 5 and Fig. 6 show the results of simulation. We see that using implicit feedback can significantly improve the effective download ratio, especially in the environment with feedback sparsity and delay. Delaying to check the content will result in slowing down the convergence rate of trust systems. To analyze the impact of various post-download behaviors on the performance, we classify normal peers into selfish, lazy, and good peers. Selfish peers always immediately delete files regardless of the quality is good or bad. Lazy peers always delay fd time slots to check the quality of files. Good peers always timely check the quality of files, delete the polluted files and continue to share the authentic files. Let x, x 2, x 3,and x 4 denote the proportion of selfish, lazy, good, and malicious peers in PeerTrust Selfish Figure 7. Comparing the effective download ratio of various peers in three trust systems with implicit feedback. the networks respectively, the population vector can be described by X = (x, x 2, x 3, x 4 ). Fig. 7 shows the performance of AVG, HALF, and PeerTrust systems with implicit feedback when X = (,,, ). We see that all trust systems based on implicit feedback can provide differential services for peers with different behaviors. Good peers are rewarded by better and better performance while Lazy peers and selfish peers will be punished by the declining performance. It is proved that trust mechanism based on implicit feedback can encourage the rational peers to behave as expected by system. From the cost of implementation, implicit feedback derives from the post-download behavior of the local peer, which can be monitored by system automatically. Considering many attacks against trust systems derived from subjective feedback, implicit feedback can improving the robustness of trust systems. V. CONCLUSION In this paper, we analyze the impact of feedback sparsity and delay on the feedback-based trust systems in the decentralized P2P network. Based on a uniform simulation platform, we compare the performances of trust systems under the idealized feedback assumption and Lazy Good

7 88 JOURNAL OF NETWORKS, VOL. 7, NO. 8, AUGUST 22 the real feedback model considering feedback sparsity and delay. The results show that feedback sparsity and delay significantly affect the effectiveness of trust systems. We propose an implicit feedback mechanism based on the retention time of the downloaded file to removing the impact of explicit feedback on trust. Using the implicit feedback, the performance of trust systems are reviewed in the same network environment. We find the implicit feedback-based trust systems can gain better performance than the explicit feedback-based trust systems. ACKNOWLEDGMENT This work was supported by the National 973 Program of China under Grant No. 2CB7343. REFERENCES [] E. Adar and B. Huberman, Free riding on gnutella, First Monday, vol. 5, pp. 2-3, 2. [2] J. Liang, R. Kumar, Y. J.Xi, and K. W. Ross, Pollution in P2P file sharing systems, IEEE Infocom 25: The Conference on Computer Communications, vols. -4, pp , 25. doi:.9/infcom [3] S. D. Kamvar, M. T. Schlosser, and H. Garcia-Molina, The eigentrust algorithm for reputation management in p2p networks, in Proc. of the 2th International Conference on World Wide Web, pp , 23. doi:.45/ [4] R. F. Zhou, K Hwang, Powertrust: A robust and scalable reputation system for trusted peer-to-peer computing, IEEE Trans. on Parallel and Distributed Systems, pp , 27. doi:.9/tpds.27.2 [5] L. Xiong and L. Liu, PeerTrust: Supporting Reputation- Based Trust for Peer-to-Peer Electronic Communities, IEEE Trans. on Knowledge and Data Eng., vol. 6, no. 7, pp , 24. doi:.9/tkde [6] G. Boella, M. Remondino, and G. Tornese, Simulating the Human Factor in Reputation Management Systems for P2P Networks an Agent Based Model, in Proc. of the Fifth International Conference on Web Information Systems and Technologies, 29. [7] U. Lee, M. Choi, J. Cho, M. Y. Sanadidi, M. Gerla, Understanding pollution dynamics in p2p file sharing, UCLA CSD Techical Report, 25. [8] L. Mekouar, Y. Iraqi, R. Boutaba, Reputation-based trust management in peer-to-peer systems: taxonomy and anatomy, Handbook of Peer-to-Peer Networking, pp , 2. doi:.7/ [9] Z. Liang and W. Shi, Analysis of ratings on trust inference in open environments, Elsevier Performance Evaluation, vol. 65, no. 2, pp , Feb 28. doi:.6/j.peva [] Z. Liang, W. Shi, Enforcing cooperative resource sharing in untrusted peer-to-peer environment, ACM Journal of Mobile Networks and Applicationsvol., no. 6, pp , 25. doi:.7/s [] Y. Wang, J. Vassileva, Trust and reputation model in peer-to-peer networks, in: Third International Conference on Peer-to-Peer Computing, P2P 3, September 23. doi:.9/ptp [2] A. Josang, R. Ismail, The beta reputation system, in Proc. of the 5th Bled Electronic Commerce Conference, Bled, Slovenia, 22. [3] Q. Y. Feng, Y. F. Dai, Lip: A lifetime and popularity based ranking approach to filter out fake files in p2p file sharing systems, in Proc. of IPTPS, Bellevue, Washington, 27. [4] N. Fotiou, G. Marias, and G. Polyzos, Information Ranking in Content-Centric Networks, in Proc. of the Future Network and Mobile, 2. [5] R. Ormandi, I. Hegedus, K. Csernai, and M. Jelasity, Towards inferring ratings from user behavior in BitTorrent communities, 9th IEEE International Workshops on Enabling Technologies: Infrastructure for Collaborative Enterprises (Wetice 2), pp , 2. doi:.9/wetice.2 [6] M. Schlosser, T. Condie, and S. kamvar, Simulating a file-sharing p2p network, in Proc. of the st Workshop on Semantics in Peer-to-Peer and Grid Computing, pp , May 23. Zhihua Huang (976-) is currently a Ph.D. candidate at the Department of Electronic Engineering of Shanghai Jiao Tong University, Shanghai, China. She received her B.S. degree in Information Engineering and M.S. degree in Communication and Information System from Xinjiang University of Urumqi City, Xinjiang, in 999 and 25, respectively. Her research interests include peer to peer network, trust and reputation systems, and Information security. Songnian Lu (943-) is a Professor and Ph.D. supervisor at the Department of Electronic Engineering of Shanghai Jiao Tong University, Shanghai, China. He received his B.S. from Shanghai Jiao Tong University in 982. His research interests include information security, distributed systems, and Operating Systems. Aixin Zhang (973-) is an Associate professor of the School of Information Security Engineering, Shanghai Jiao Tong University, Shanghai, China. She received her Ph.D. degree in control theory and control engineering From East China University of Science and Technology, Shanghai, China, in 23, her M.S. degree in control theory and control engineering and B.E. degree in computer software engineering from Liaoning Shihua University, Liaoning, China, in 998 and 995, respectively. Her current research interests include information security, digital rights management and multimedia information forensics. Jie Gu (982-) is currently a Ph.D. candidate at the Department of Electronic Engineering of Shanghai Jiao Tong University, Shanghai, China. He received his B.S. and M.S. degrees in Telecommunication Engineering from Xidian University in 24 and 27, respectively. His research interests include anonymous authentication, key distribution, wireless network security and information security.