Page 72 - Trust in ICT 2017
P. 72
1 Trust in ICT
BRS uses a simple discounting procedure for handling false opinions. The discounting is based on the level of
trust the BRS places in the agents that provide opinions. For instance, if BRS considers an agent to be very
untrustworthy as a service provider, it heavily discounts its opinions. Such assumption is sometimes called
trust transitivity, because it states that if an agent is trustworthy to provide a certain service it can also be
trusted to provide good (honest) opinions.
8.2.1.3 Abdul-Rahman, Hailes (ARH)
The trust model proposed by Abdul-Rahman and Hailes (ARH) [68] uses qualitative information for computing
and representing trust. In ARH, domains of trust degrees and assessments are the same: X=K={vb < b < g <
vg}, where elements denote ‘very bad’, ‘bad’, ‘good’, and ‘very good’ degrees (assessments), respectively.
ARH copes with liars by using a mechanism capable of correcting opinions. For instance, ARH can learn if an
agent consistently badmouths other agents and adjusts its opinions accordingly. Additionally, ARH is the only
tested trust model that separates trust by service types.
8.2.1.4 Travos (TRA) [69]
Travos (TRA) is a trust and reputation model for agent-based virtual organizations. Similar to BRS it is based
on the beta distribution and represents trust degrees as its expected value. Moreover, feedback in Travos is
also represented in the form of 2-tuples<m, n>, but contrary to BRS, Travos uses binary interaction outcomes.
Thus (1, 0) represents a satisfactory and (0, 1) an unsatisfactory interaction. The interpretation component
computes these tuples by thresholding the interaction outcomes; if the outcome reaches the threshold, we
get (1, 0), if not, (0, 1). Like ARH, there are three thresholds; TRAL thresholds at 0.25, TRAM at 0.50, and TRAH
at 0.75.
Travos expects opinions as tuples hr, si that contain the number of positive, r, and negative, s, past
interactions. When a receives an opinion, say (ai, aj, s, t, 0.60, 0.05), the interpretation component simulates
a number of interactions of ai with aj by using truncated normal distribution. It sets the mean to the opinion’s
internal trust degree, 0.60, and the standard deviation to the same value that is used for generating
experiences, 0.10. Each sampled number is then compared against the threshold to determine whether the
interaction is satisfactory. This procedure assures that a obtains the same tuple – adjusted for the correctness
of the given opinion – that would have been obtained if agent ai had interacted with aj 10 times and then
reported the number of positive and negative interactions. For instance, with threshold 0.50, the opinion
above would most likely be transformed into hai, aj, s, t, h8, 2i, 0.05i.
Travos computes confidence in its experiences and if confidence is not sufficient, it combines experiences
with opinions. Additionally, it also uses a complex mechanism to reduce the effect of false opinions. If an
opinion provider is deemed as a liar, Travos reduces the weight of its opinions. Travos manipulates
parameters of the beta distribution.
8.2.1.4 Eigen Trust [70]
EigenTrust is a trust model for P2P networks. It computes global trust values based on opinions from all peers
in the system. An important aspect of EigenTrust is the notion of special peers that are pre-trusted. The trust
in those peers has to be accurate, otherwise EigenTrust’s computation method does not converge. EigenTrust
paper does not specify how to determine such peers.
EigenTrust uses binary interaction outcomes and computes local trust values in the form of net difference
between the number of positive and negative interactions. If the difference is negative - more negative than
positive interactions - EigenTrust assigns a local trust value of 0 to such peer. Because of this, it is said that
EigenTrust does not measure negative trust, since it cannot differentiate between peers with whom it has
had bad experiences from those with whom it has not interacted.
EigenTrust also exchanges opinions in the form of tuples that contain the number of positive and negative
past interactions. EigenTrust does not have any special mechanism to deal with false opinions. Similar to
BRS, it considers trust to be transitive, and simply discounts opinions based on the level of trust it has in
agents as service providers.
64