Research On Trust Model Based On Recommendation Evidence In Unstructured P2P Network

In recent years, P2P network grows rapidly. P2P network solves problems, such asover-reliance on server in traditional client/server(C/S) architecture. However, P2Pitself has the characteristics of self-organization, anonymity and openness, whichprovides a way for the spread of illegal files and virus. Trust model is a reliable wayto solve the security issues of P2P network. This paper is about the trust model inunstructured P2P network. We emphatically discuss the trust model based onrecommendation evidence in this paper. The methods to identify dishonestrecommendation in traditional trust model have certain limitations, such as the largedeviation in certain environments. Thus, we propose an improved trust model whichis based on recommendation evidence. For the problem of identification to dishonestyrecommendation, this paper set the process of recommendation as an interactiveprocess. So the evidence of trust, which calculated by our model, not only on behalfof the credibility of the services but also, represents the reliability of recommendation.This paper proposed a new method for calculation of trust uncertainty. In theprocess of introducing D-S evidence theory to trust model which is based onrecommendation. The main point is to establish the basic probability assignmentfunction. The uncertainty of evidence represents the unknown level from source peerto destination peer. Because of this, the uncertainty can guide the behavior of peers.And it can be the dynamic factor in the process of combining local trust andrecommendation from network. This paper presents a method for calculating theuncertainty which is combined the impact of collision rate and the number ofinteractions.In the calculation of trust, this paper presents a new decay function, which usedifferent treatments to positive records and negative records. The fast decay ofpositive records can encourage peers providing services more actively; simultaneouslyslowly decay of negative records can punish misconduct of node. We also add analgorithm to detect dynamic node in our model. This algorithm can compare historicalevidence, current evidence and calculate the conflict factors to determine whether thebehavior of peer has been suddenly changed.Recommendation is weighted by overall trust before synthesis of recommendationbecause the trust of our model can represent the reliability of the recommendation. When combining recommendation from network and local evidence, our model use adynamic factor to weight. The results of synthesis using original D-S evidence theoryare not ideal when the conflict between evidence is large. Taking into account theusing environment of our model, this paper adopt an improved D-S evidencecombination theory and this theory has been shown more reasonable in distributedenvironment.For the recommended searching problem, the overall strategy of searching in ourmodel is to use Gossip algorithm and probability mechanism. So peers only route thequery message to their trustful neighbors, which reduce the probability of receivingthe trust request message on malicious peers. Compared to the traditional flood searchalgorithm, our algorithm greatly reduces the flux of network and improves efficiencyof network.At last, to evaluate our model, we conduct a series of comparative experiments withEigenRep, PeerTrust and TBRM. In the different modes of attack, we compare thecurve of successful transaction rate of four kinds of models and do a detailed analysisof the experimental results. The results show that our algorithm has improved thesuccessful transaction rate and can better adapt the large-scale distributed P2Pnetwork. This paper also compares the Gossip algorithm based on the probability andtraditional flood search algorithm. According to the statistical results we draw curveof recall rate, precision rate and successful transaction rate. Experimental results showthat Gossip algorithm plays an important active role in overall search strategy.