FUTURE TRUST FORECAST IN OPEN MOBILE AGENT
ENVIRONMENT
Kłopotek A. Mieczysław
Institute of Computer Science, Polish Academy of Sciences, ul. J.K. Ordona 21, 01-237 Warszawa, Poland
Wolski Michał
Institute of Computer Science, University of Podlasie, ul. Sienkiewicza 51, 08-110 Siedlce, Poland
Keywords: Mobile agents, open environment reputation, simulation, trust.
Abstract: Open Mobile Agent Environments, where entities can join and leave the system at any time, are particularly
susceptible to the attacks of malicious entities. Hence intense studies of potential solutions of related
problems are needed, including proper and speedy estimation of node’s trustworthiness. We propose an
optimisation method for trust estimation (reputation forecasting) and apply it to two known reputation
metrics (eBay and BetaSystem). We show results of simulations comparing the effectiveness of reputation
discovery using the original algorithm and the optimized (forecasting reputation) one.
1 INTRODUCTION
Reputation mechanisms allow agents to establish
trust in other agents' intentions and capabilities in
the absence of direct interactions. In the context of
e-commerce, the parties involved in mutual
interactions may publicly rate their trading partner in
terms of his compliance to the terms of trade (e.g. on
eBay or Yahoo! Auctions). This benefits other, new
agents considering interacting with those partners,
who would otherwise have no idea about their
trustworthiness. Reputation systems are an important
building block for achieving trust within large
distributed communities, especially when mutually
unknown agents engage in ad-hoc transactions. In
this article we present design of Open Mobile
Agent’s Environment Simulator and simulation
results obtained with that tool. To demonstrate
simulator features, we compare by simulation known
reputations metrics (eBay and BetaSystem)
algorithms and a new trust optimization algorithm
(FutureTrust) that is dedicated especially to Open
Mobile Agent’s Environment, with respect to their
efficiency in identifying trustworthiness of nodes.
2 OPEN ENVIRMONEMT ISSUES
Expansion of mobile agents software is due to the
business requirements that need software which will
co-operate with each other without early
coordination. Hence agent has to have a trust to
other agents before he begins transaction.
By trust we (or symmetrically, distrust) mean "... a
particular level of the subjective probability with
which an agent will perform a particular action, both
before he can monitor such action (or independently
of his capacity to monitor it) and in a context in
which it affects his own action.”(Misztal, 2004).
In large-scale open distributed systems, trust remains
a fundamental challenge for the success of their
operation. When we use Open Mobile Agent’s
Environment we think about scenarios like:
system open in that agents can enter and
leave at any time. This means that an agent
could change its identity on re-entering and
avoid punishment for any past wrongdoing.
system that allows agents with different
characteristics (for example, policies,
abilities, roles) to enter it and interact with
each other (Ramchurn and Jennings, 2004)
353
A. Mieczysław K. and Michał W. (2007).
FUTURE TRUST FORECAST IN OPEN MOBILE AGENT ENVIRONMENT.
In Proceedings of the Third International Conference on Web Information Systems and Technologies - Internet Technology, pages 353-356
DOI: 10.5220/0001271003530356
Copyright
c
SciTePress
no agent can know everything about its
environment (Huynh and Jennings, 2006).
Co-operation without early coordination implies that
agent has to have a trust to other agents and to the
infrastructure before he begins transaction.
During his journey agent can interact with each of
nodes and learn their behaviour over a number of
encounters. This knowledge has to be memorized.
The agent faces challenges like:
unknown network topology,
evil (hostile) nodes damaging agents
An agent cannot cope with these issues alone,
communities have to be formed. To simplify the
problem, we assume that families of agents (agents
that can fully trust one another, if meeting on a
trusted node and sharing a common repository on
trusted nodes) are sent out into an open environment.
The problems of mobile agent trust and security in
open environment are of extreme complexity. In this
part of our research we skip communication problem
between two or more agents and between agent and
common repository. At this moment let us devote
our attention only to reputation metrics.
While trusting one another, the family members
have to evaluate appropriately the trust they may
have to the environment. A number of potentially
suitable global reputation systems, such as eBay,
BetaSystem (Jøsang and Ismail, 2002), and local
ones like EigenTrust (Kamvar and Schlosser, 2003),
Sporas (Zacharia and Maes, 2000) have been
elaborated, while other are under development.
Comparative studies of usefulness of these metrics
are needed and some tools have been elaborated for
such analyses. Complex comparison of various
metrics can be found in (Schlosser and Andreas,
2005). However, the testbed for metrics presented
there is not suitable for our purposes of study of
open environments, hence we built a Mobile Agents
Reputation Simulator (MARS) which is a useful tool
to compare global reputation systems (Wolski and
Klopotek, 2006).
3 SIMULATIONS
In our research we test the effectiveness of trust
algorithms by simulating an environment adhering
to some predefined model, which is unknown for
agents. Agents move from one node to another.
During his journey an agent interacts with nodes and
learns their behaviour over a number of encounters.
Knowledge about node behavior will have to be
stored in common repository, in form of a
“reputation level”, which is then compared to the
“intrinsic” one (the one from the predefined
simulation model).
In this paper we investigate with our simulator two
of them: eBay Algorithm and BetaSystem
Algorithm, which will be subject to our optimization
(FutureTrust). Each algorithm has been tested on
the same network, created in a random way, with
topological features similar to the Internet.
We investigated networks consisting of :
good nodes, which have attractive
information for agents,
neutral nodes, which have nothing
interesting for agents,
evil (bad, hostile) nodes, which destroy
agent in case of interaction between agent
and node.
We experimented with four node groups: good
node family, neutral node family, and a random
node family, composed of a mixture of nodes
described in the previous three groups. Last one is
variable node family, which consisted nodes which
are very unstable. They change dynamically their
behavior to towards visiting agents with each
encounter. Behavior of “variable” nodes is based on
normal distribution and is random with probability
equal 0.33 for each kind of behavior.
3.1 BetaSystem and eBay Algorithm
First trust metrics, that we investigated, is well
known eBay algorithm, which needs to maintain
information on good and all transactions.
The next one was the Bayesian Reputation System
called BetaSystem, allowing each agent to rate node
positively or negatively (Jøsang and Ismail, 2002).
In our simulations positive ratting is given to good
nodes, and negative ratting is obtained by neutral
and bad nodes.
3.2 FutureTrust Algorithm
Our agents have common repository, that means if
one of agents has a good transaction each agent of a
family will know about it. If a second agent has
good transaction with particular nodes we can
forecast that next agent will good transaction too. If
so, we can construct a metric exploiting foreseen
trust values in some iterations in the future.
In our research we consolidate known reputation
metrics with stochastic process, which can tell us
forecast reputation with particular probability in
defined time in future. We sought to minimize risk
relevant with forecasting of trust value and we want
to answer to question: “What trust value will have
WEBIST 2007 - International Conference on Web Information Systems and Technologies
354
this family of node’s in the next iteration, next 10,
100 iterations?” To come to a solution we use two
simplifying assumptions:
trust value is similar to random walk, with
reputation in short time period being a
random variable with normal distribution
for any time in the future reputation has
log-normal distribution character.
Based on log-normal distribution we can forecast
future value of trust (equation 1)
],)[ln()ln( TTRR
T
μμφ
+
(1)
where: T – time (number of iterations)
R – present reputations (enumerated by known trust
metrics)
R
T
– future reputation (in T iterations)
μ – variable responsibility for fluctuation of trust
metrics
To compute future trust value we have to store
information about positive and negative transactions
in an incremental table.
Based on equation 1 we compute FutureTrust as
(equation 2)
Where c – value of standard deviations
(e.g. if μ=95% then c=1.96 )
(2)
3.3 Comparison of Algorithms
Subsequent figures are representative to all
experiments. We can see, that agents need time to
learn node family trust estimate. For the small
network we used (about 1000 nodes), the number of
iterations (equal to the number of transactions of
each surviving agent) needed by any algorithm was
at most 30.
0 1 2 3 4 5 6 7 8 9 1
0
1
1
1
2
1
3
1
4
1
5
1
6
1
7
1
8
1
9
2
0
2
1
2
2
2
3
2
4
2
5
2
6
2
7
2
8
2
9
3
0
0,000
0,200
0,400
0,600
0,800
1,000
Algorithm comparison on good node family
BetaSystem vs. Ebay vs. FutureTrust
BetaSystem FutureTrust eBay
time (iterations
)
trust value
We assume that the best reputation algorithm is
that one, which allows to set faster the correct value
of trust for particular family of nodes.
On next figures we present a comparison
between the basic algorithm version and the
FutureTrust modifier. The FutureTrust parameters
were set to: T = 50 (iterations) and μ = 0.01 - 1%
changes of reputation.
First comparison refers to good nodes family.
Figure 1 shows that eBay algorithm is very fast to
recognize true reputation of node family.
0 1 2 3 4 5 6 7 8 9 1
0
1
1
1
2
1
3
1
4
1
5
1
6
1
7
1
8
1
9
2
0
2
1
2
2
2
3
2
4
2
5
2
6
2
7
2
8
2
9
3
0
0,000
0,050
0,100
0,150
0,200
0,250
0,300
0,350
0,400
0,450
0,500
Algorithm comparison on neutral node family
BetaSystem vs. Ebay vs. FutureTrust
BetaSystem FutureTrust eBay
time (iterations
)
trust value
Figure 2 presents results for the second group of
nodes: neutral one. We think that in that case the
best algorithm is FutureTrust algorithm, because it is
growing up to correct value of trust. Instead eBay
algorithm is worthless because it doesn’t notice any
value.
Next family of nodes is the random family. This
family is built of three types of nodes: good, neutral,
evil and have constant structure of behaviour. It
means node never changed their behaviour to any
agents.
0 1 2 3 4 5 6 7 8 9 1
0
1
1
1
2
1
3
1
4
1
5
1
6
1
7
1
8
1
9
2
0
2
1
2
2
2
3
2
4
2
5
2
6
2
7
2
8
2
9
3
0
0,000
0,050
0,100
0,150
0,200
0,250
0,300
0,350
0,400
0,450
0,500
0,550
0,600
Algorithm comparison on random node family
BetaSystem vs. Ebay vs. FutureTrust
BetaSystem FutureTrust eBay
time (iterations
)
trust value
>=
<=
+
++
goodbadforeR
goodbadforeR
TcTR
T
TcTR
T
μμ
μμ
)ln(
)ln(
Figure 1: Algorithm comparison for good node family.
Figure 2: Algorithm comparison for neutral node family.
Figure 3: Algorithm comparison for random node family.
FUTURE TRUST FORECAST IN OPEN MOBILE AGENT ENVIRONMENT
355
In that case the best algorithm is FutureTrust
because its value of reputation for this kind of family
is set to correct value faster than other algorithms.
Last but not least node family is the variable node
family. It is very similar to random node family but
the main difference is that in variable family each
node always changes its behavior to agents.
In that part of our research we make assumption that
three types of behavior of nodes (good, neutral, evil)
switch with probability equal 0.333.
0 1 2 3 4 5 6 7 8 9 1
0
1
1
1
2
1
3
1
4
1
5
1
6
1
7
1
8
1
9
2
0
2
1
2
2
2
3
2
4
2
5
2
6
2
7
2
8
2
9
3
0
0,000
0,050
0,100
0,150
0,200
0,250
0,300
0,350
0,400
0,450
0,500
Algorithm comparison on variable node family
BetaSystem vs. Ebay vs. FutureTrust
BetaSystem FutureTrust eBay
time (iterations
)
trust value
Figures 4 demonstrates that in that case the best
algorithm is BetaSystem or Future Trust algorithm,
because only these algorithms set correct value of
trust for the family.
Above figures show that assignment of forecast
value of trust based on known metrics (equation 2)
allows to reduce the number of iterations, which are
needed to correctly recognize true reputation of
nodes family.
It means, if we are forecasting future value of
trust we can get benefits such us:
- faster recognition of true reputation of nodes,
- less cost of agents function,
- less load of agents system,
- less consumption of memory, where we store
information about network.
4 CONCLUSION
In this paper we investigated some problems
encountered when computing reputation in open
environment. We reported on a comparative study
two known metrics eBay and BetaSystem and our
own based on future trust optimization.
We pointed at the very important problem of
speed of recognition of intrinsic reputation for a
family of nodes and demonstrated that our
innovative technique based on forecasting trust
value offers a solution.
We showed that usage of the FutureTrust
optimization formula can reduce significantly the
cost of computations related to trust determination.
While the current paper concentrates on trust
estimation, a more important issue is to device a
mechanism that allows exploration of only most
trusted part of network so that agents can collect
information (resources) faster. Beside this, in our
future research we will check what happens when
some nodes will clone or change a mobile agent and
how different ways of mobile agents interaction with
the common repository will have influence on
reputation value in particular node family.
REFERENCES
Huynh D., Jennings N.R., Shadbolt N.R.: An integrated
trust and reputation model for open multi-agent
systems, Autonomous Agents and Multi-Agent
Systems, Issue: Volume 13, Number 2, September
2006
Jøsang A. and Ismail R. The Beta Reputation System. In
Proceedings of the 15th Bled Conference on
Electronic Commerce, Bled, Slovenia, 17-19 June
2002,
Kamvar S. D., Schlosser M. T., Garcia-Molina H. :The
Eigentrust Algorithm for Reputation Management in
P2P Networks, (http://dbpubs.stanford.edu), 2003
Kłopotek M. A., Wolski M.: Comparative Study of Trust
Algorithms for Mobile Agent in Open Environment,
Proceedings of Artificial Intelligence Studies Vol.3,
(26)/2006, pp 213-220
Misztal, B.: Trust in Modern Societies, Polity Press,
Cambridge, Mass., (1996)
Ramchurn S.D., Jennings N. R.: Trust in agent-based
software, Cyber Trust & Crime Prevention Project,
2004
Schlosser, Andreas, Voss, Marco and Brückner, On the
Simulation of Global Reputation Systems. Journal of
Artificial Societies and Social Simulation, Lars 2005
Wolski M. Kłopotek M.A.: A Concept of Reputation for
Mobile Agents Environments, chapter in Polish Journal
of Environmental Studies Vol. 15, No. 4C, 2006, pp.
207-211, ISSN 1230-1485, Świnoujście 2006
Zacharia G. and Maes P.: Trust Management through
Reputation Mechanisms. Applied Artificial
Intelligence, 14, 2000.
Figure 4: Algorithm comparison for variable node family.
WEBIST 2007 - International Conference on Web Information Systems and Technologies
356