Trust - Peer-to-Peer Computing: Applications, Architecture, Protocols, and Challenges - page 147

the total number of aggregators in some cluster) that are trying to cheat
the system by submitting erroneous aggregation results. Now, the global
trust value of v is updated by v's aggregators without decrypting the
intermediate reports, thereby ensuring privacy and integrity of votes.

Simulation results indicated that SuperTrust performed well in terms of
message overhead and response time.

6.8 PowerTrust
Zhou and Hwang proposed a DHT-based trust management scheme called
PowerTrust [Zhou and Hwang, 2007b]. Figure 6.12(a) shows the system archi-
tecture of PowerTrust. First, a trust overlay network (TON) is built on top
of all peers in the P2P system. All peers send local trust scores among them-
selves periodically. The PowerTrust system then aggregates the local scores
to compute the global reputation score of each participating peer. All global
scores form a reputation vector V = (v 1 , . . . , v n ), where

i v i = 1.
As shown in Figure 6.12(a), the regular random walk module supports the
initial reputation aggregation. The lookahead random walk (LRW) module is
used to update the reputation score periodically. To this end, the LRW also
works with a distributed ranking module to identify the power peers (i.e.,
peers with prior trust). The system leverages the power peers to update the
global reputation scores.
In PowerTrust, feedback scores are generated by Bayesian learning or by
an average rating based on peer satisfaction. Each peer normalizes all issued
feedback scores. Consider the trust matrix R = (r ij ) defined over the n-peer
TON, where r ij is the normalized local trust score defined by r ij =

s ij
P

j s ij ,
and s ij is the most recent feedback score that peer i rates peer j. If there is
no link from peer i to peer j, s ij is set to 0. Thus, for all 1≤i, j≤n, we have
0≤r ij ≤1 and∀i,

n
j=1 r ij = 1. In other words, matrix R is a stochastic
matrix in which all entries are fractions and each row sums to 1. This requires
that the scores issued by the same peer to other peers are normalized.
The reputation vector is then recursively updated as follows:

V (t+1) = R T ×V (t)

(6.13)

This updating is done until|V (i) −V (i−1) |≤ǫ.
Using the parameter sweeping P2P Grid applications in their simulation
study, Zhou and Hwang found that PowerTrust performed well in terms of
trust aggregation accuracy, message overhead, and makespan of the Grid ap-
plications.