Trust networks

Trust Networks

Krishnaprasad Thirunarayan (T. K. Prasad)
Professor, Department of Computer Science and Engineering
Kno.e.sis - Ohio Center of Excellence in Knowledge-enabled Computing
Wright State University, Dayton, OH-45435
(Collaborators: Pramod Anantharam, Cory Henson, and Professor Amit Sheth)

CTS-2012

May 21, 2012 Trust Networks: T. K. Prasad 1

Broad Outline
• Real-life Motivational Examples (Why?)
• Trust : Characteristics and Related Concepts (What?)
• Trust Ontology (What?)
– Type, Value, Process, Scope
• Gleaning Trustworthiness (How?) + Robustness to Attack
– Practical Examples of Trust Metrics
– Comparative Analysis of Bayesian Approaches to Trust
• Research Challenges (Why-What-How?)
– Sensor Networks
– Social Networks
– Interpersonal
• Details of Bayesian Approach to Multi-level Trust

Real-life Motivational Examples

(Why track trust?)


Interpersonal

• With which neighbor should we leave our
children over the weekend when we are
required to be at the hospital?

• Who should be named as a guardian for our
children in the Will?


Social
• In Email:

– SUBJECT: [TitanPad] Amit Sheth invited you to an
EtherPad document.
– CONTENT: View it here:
http://knoesis.titanpad.com/200

• Issue: Is the request genuine or a trap?


Social

• To click or not to click a http://bit.ly-URL

• To rely or not to rely on a product review
(when only a few reviews are present, or the
reviews are conflicting)?


Sensors
• Weather sensor network predicts a potential tornado in
the vicinity of a city.
• Issue: Should we mobilize emergency response teams
ahead of time?
• Van’s TCS (Traction Control System) indicator light came
on intermittently, while driving.
• Issue: Which is faulty: the indicator light or the traction
control system?
• Van’s Check Engine light came on, while driving.
• Issue: Which is faulty: the indicator light or the
transmission?

Man-Machine Hybrid
Collaborative Systems
The 2002 Uberlingen Mid-air Collision (between
Bashkirian Airlines Flight 2937 and DHL Flight
611) occurred because the pilot of one of the
planes trusted the human air traffic controller
(who was ill-informed about the unfolding
situation), instead of the electronic TCAS system
(which was providing conflicting but correct
course of action to avoid collision).
http://en.wikipedia.org/wiki/2002_Uberlingen_mid-air_collision


Man-Machine Hybrid
Collaborative Systems
In hybrid situations, artificial agents
should reason about the
trustworthiness and deceptive
actions of their human counter
parts. People and agents in virtual
communities will deceive, and will
be deceived. Castelfranchi and Tan, 2002


Common Issues and Context
• Uncertainty
– About the validity of a claim or
assumption
– Past Experience : Vulnerability

• Need for action
• Critical decision with potential for loss


Commonality among Trust Definitions*
• a Trustor
– someone who must choose whether, and how much, to
trust
• a Trustee
– someone or something that is to be trusted
• an Action
– by which the trustor is choosing to be vulnerable to the
trustee based on an assessment of trustee’s nature
• a Context
– in which the potential negative consequences of betrayal
outweigh any perceived positive results.
*http://www.iarpa.gov/rfi_trust.html

Sources of uncertainty
• In social network, trustor’s incomplete
knowledge or trustee’s devious intentions.
• In sensor network, unpredictable environment
(e.g., random phenomenon), or sensor faults
(e.g., due to aging), leading to corrupt data.
• In cognitive radio networks, unpredictable
channel noise or traffic.
• Interpersonal examples
– Irresponsible / selfish / greedy / vengeful traits


Why Track Trust?
• In Mobile Ad Hoc Networks (MANETs),
trust enables dynamic determination of
secure routes.
– Efficiency: To improve throughput
• By avoiding nodes facing bad channel condition
– Robustness : To detect malicious nodes
• When attackers enter the network in spite of
secure key distribution/authentication


Why Track Trust?
• In sensor networks, it allows detection of
faults and transient bad behaviors due to
environmental effects.

• In cognitive radio networks, it can enable
selection of optimal channel (less noisy,
less crowded channels).


Why Track Trust?
• In E-commerce:

–To predict future behavior in a
reliable manner.
–To incentivize “good” behavior and
discourage “bad” behavior.
–To detect malicious entities.


The Two Sides of Trust
• Trustor assesses trustee for
dependability.
• Trustee casts itself in positive light to
trustor.
• Trust is a function of trustee's
perceived trustworthiness and the
trustor's propensity to trust.

Risk/uncertainty mitigation
• Compensating factors that alter trust
threshold
• In e-commerce, warranties and insurance
reduce risk.
• In sensor networks, redundancy enables
filtering of corrupt data.
• In interpersonal situations, close ties
help.


Trust and Related Concepts

(What is trust?)


Trust Definition : Psychology slant

Trust is the psychological state
comprising a willingness to be
vulnerable in expectation of a
valued result.

Ontology of Trust, Huang and Fox, 2006
Josang et al’s Decision Trust


Trust Definition : Psychology slant

Trust in a person is a commitment to
an action based on a belief that the
future actions of that person will
lead to good outcome.

Golbeck and Hendler, 2006


Trust Definition : Probability slant
Trust (or, symmetrically, distrust)
is a level of subjective probability
with which an agent assesses
that another agent will perform
a particular action, both before
and independently of such an
action being monitored …
Can we Trust Trust?, Diego Gambetta, 2000
Josang et al’s Reliability Trust

Trustworthiness Definition :
Psychology Slant
Trustworthiness is a collection of
qualities of an agent that leads them
to be considered as deserving of
trust from others (in one or more
environments, under different
conditions, and to different degrees).
http://www.iarpa.gov/rfi_trust.html

Trustworthiness Definition :
Probability slant
Trustworthiness is the objective
probability that the trustee
performs a particular action on
which the interests of the
trustor depend.
Solhaug et al, 2007


Trust vs Trustworthiness : My View
Trust Disposition
Depends on
Potentially Quantified Trustworthiness Qualities
+
Context-based Trust Threshold
E.g.*, In the context of trusting strangers, people in
the West will trust for lower levels of trustworthiness
than people in the Gulf. *Bohnet et al, 5/2010


Qualitative Trust Dispositions
(depends on Trust Threshold)

http://paulenglish.com/trust.html


(Community-based) Reputation
• Reputation* is the community or public
estimation of standing for merit,
achievement, reliability, etc. *dictionary.com
• Reputation** is the opinion (or a social
evaluation) of a community toward a
person, a group of people, or an
organization on a certain criterion. **Wikipedia
• Cf. Brand-value, PageRank, eBay profile, etc.


Trust vs. (Community-based)
Reputation
Reputation can be a basis for trust.
However, they are different notions*.

• I trust you because of your good reputation.
• I trust you despite your bad reputation.
• Do you still trust Toyota brand?

*Josang et al, 2007


Trust vs. (Community-based)
Reputation
Trust :: Reputation
::::
Local :: Global
::::
Subjective :: Objective
(Cf. Security refers to resistance to attacks.)


Reputation is Overloaded

Community-based Reputation
vs.
Temporal Reputation

(Cf. Sustained good behavior over time elicits
temporal reputation-based trust.)


Trust is well-known,
but is not well-understood.
The utility of a notion
testifies not to its clarity but
rather to the philosophical
importance of clarifying it.
-- Nelson Goodman
(Fact, Fiction and Forecast, 1955)


Trust Ontology

(What is trust?)

Illustration of Knowledge Representation and Reasoning:
Relating Semantics to Data Structures and Algorithms


Example Trust Network -
Different Trust Links with Local Order on out-links

• Alice trusts Bob for recommending good car
mechanic.
• Bob trusts Dick to be a good car mechanic.
• Charlie does not trust Dick to be a good car
mechanic.
• Alice trusts Bob more than Charlie, for
recommending good car mechanic.
• Alice trusts Charlie more than Bob, for
recommending good baby sitter.
*Thirunarayan et al, IICAI 2009

Digression: Illustration of Knowledge
Representation and Reasoning
• Abstract and encode clearly delineated “subarea”
of knowledge in a formal language.
– Trust Networks => node-labeled, edge-labeled
directed graph (DATA STRUCTURES)
• Specify the meaning in terms of how “network
elements” relate to or compose with each other.
– Semantics of Trust, Trust Metrics => using logic or
probabilistic basis, constraints, etc. (SEMANTICS)
• Develop efficient graph-based procedures
– Trust value determination/querying (INFERENCE
ALGORITHMS)


(In recommendations)

(For capacity to act)

(For lack of
capacity to act)

Trust Ontology*
6-tuple representing a trust relationship:
{type, value, scope, process}
trustor trustee

Type – Represents the nature of trust relationship.
Value – Quantifies trustworthiness for comparison.
Scope – Represents applicable context for trust.
Process – Represents the method by which the value is
created and maintained.
*Anantharam et al, NAECON 2010

Trust Ontology:
Trust Type, Trust Value, and Trust Scope

 Trust Type*
 Referral Trust – Agent a1 trusts agent a2’s ability to
recommend another agent.
 (Non-)Functional Trust – Agent a1 (dis)trusts agent a2’s
ability to perform an action.
 Cf. ** trust in belief vs. trust in performance
 Trust Value
 E.g., Star rating, numeric rating, or partial ordering.
 Trust Scope*
 E.g., Car Mechanic context. *Thirunarayan et al, IICAI 2009
** Huang and Fox, 2006

Multidimensional / Orthogonal
Trust Scopes in Ecommerce

• Trust in a vendor to deliver on
commitments
• Trust in vendor's ethical use of
consumer data
• Trust in Internet communication being
secure.
• Plus: Propensity/Disposition to trust

Trust Ontology:
Trust Process
 Represents the method by which the value
is computed and maintained.
 Primitive (for functional and referral links)*
 Reputation – based on past behavior (temporal) or
community opinion.
 Policy – based on explicitly stated constraints.
 Evidence – based on seeking/verifying evidence.
 Provenance – based on lineage information.
 Composite (for admissible paths)**
 Propagation (Chaining and Aggregation)
*Anantharam et al, NAECON 2010
**Thirunarayan et al, IICAI 2009

Trust Ontology


Example Trust Network illustrating Ontology Concepts

Dick is a
Bob is a car certified
aficionado mechanic

type: referral
process: reputation
scope: car mechanic
value: TAB Bob type: functional
process: policy
TAB > TAC scope: car mechanic ASE certified
value: 10

type: referral
Alice
process: reputation
scope: car mechanic type: non-functional
value: TAC process: reputation
scope: car mechanic Dick
value: 3

47
Charlie

Unified Illustration of Trust Processes
Scenario : Hiring Web Search Engineer - An R&D Position
Various Trust Processes :
• (Temporal-based) Reputation: Past job experiences
• (Community-based) Reputation: Multiple
references
• Policy-based: Score cutoffs on screening test
• Provenance-based: Department/University of
graduation
• Evidence-based: Multiple interviews (phone, on-
site, R&D team)


Deception
• Deception is the betrayal of trust.

–Ironically, trust makes us prone to
deception.
–Knowing what features are used to
glean trustworthiness can also assist
in avoiding detection while deceiving.

Gleaning Trustworthiness :
Practical Examples
(How to determine trustworthiness?)


Trust Metric and Trust Model

• Trust Metric => How is primitive trust
represented?
• E.g., Real number, Finite levels, Partial Order, Opinion =
(belief, disbelief, uncertainty), etc.

• Trust Model => How is composite trust
computed or propagated?
Y. L. Sun, et al, 2/2008

Ideal Approach
• Capture semantics of trust using
– axioms for trust propagation, or
– catalog of example trust networks that are
equivalent.

• Develop trust computation rules for
propagation (that is, chaining and
aggregation) that satisfy the axioms or
equivalence relation.


Direct Trust : Functional and Referral
Reputation-based Process
(Using large number of observations)


Reputation-based Frameworks

• Centralized Trust Authority
– E.g., E-commerce systems, etc.

• Distributed Trust Representation and
Computation (using Bayesian analytics)
– E.g., MANETs, peer-to-peer networks, etc.


Using Large Number of Observations

• Over time (<= Referral + Functional) :
Temporal Reputation-based Process
– Mobile Ad-Hoc Networks
– Sensor Networks
• Quantitative information
(Numeric data)
• Over agents (<= Referral + Functional) :
Community Reputation-based Process
– Product Rating Systems
• Quantitative + Qualitative information
(Numeric + text data)


Desiderata for Trustworthiness
Computation Function
• Initialization Problem : How do we get initial value?
• Update Problem : How do we reflect the observed
behavior in the current value dynamically?
• Trusting Trust* Issue: How do we mirror uncertainty
in our estimates as a function of observations?
• Law of Large Numbers: The average of the results obtained from a
large number of trials should be close to the expected value.
• Efficiency Problem : How do we store and update
values efficiently?
*Ken Thompson’s Turing Award Lecture: “Reflections on Trusting Trust”


Mathematical Background

Beta PDF for Reputation


Beta-distribution : Gently
• Consider a (potentially unfair) coin that comes up
HEADS with probability p and TAILS with probability
(1 – p).
• Suppose we perform ( r + s ) coin tosses and the coin
turns up with HEADS r times and with TAILS s times.
• What is the best estimate of the distribution of the
probability p given these observations?
=> Beta-distribution with parameters ( r+1, s+1 )
f(p; r+1, s+1)


Beta Probability Density Function(PDF)

x is a probability,
so it ranges from 0-1

If the prior distribution of x is
uniform, then the beta
distribution gives posterior
distribution of x after
observing a-1 occurrences
of event with probability x
and b-1 occurrences of the
complementary event with
probability (1-x).


a = b, so the pdf’s are symmetric w.r.t 0.5.
Note that the graphs get narrower as (a+b) increases.

a= 2
a= 1
b= 2
b= 1

a= 5 a= 10
b= 5 b= 10


a ≠ b, so the pdf’s are asymmetric w.r.t . 0.5.
Note that the graphs get narrower as (a+b) increases.

a= 5 a= 5
b= 25 b= 10

a= 25 a= 10
b= 5 b= 5


Beta-distribution - Applicability

• Dynamic trustworthiness can be
characterized using beta probability
distribution function gleaned from total
number of correct (supportive) r = (a-1)
and total number of erroneous
(opposing) s = (b-1) observations so far.
• Overall trustworthiness (reputation) is its
mean: a/a +b

Why Beta-distribution?
• Intuitively satisfactory, Mathematically precise, and
Computationally tractable
• Initialization Problem : Assumes that all probability values
are equally likely.
• Update Problem : Updates (a, b) by incrementing a for
every correct (supportive) observation and b for every
erroneous (opposing) observation.
• Trusting Trust Issue: The graph peaks around the mean, and
the variance diminishes as the number of observations
increase, if the agent is well-behaved.
• Efficiency Problem: Only two numbers stored/updated.


Information Theoretic Interpretation
of Trustworthiness Probability

• Intuitively, probability values of 0 and 1 imply
certainty, while probability value of 0.5 implies
a lot of uncertainty.

• This can be formalized by mapping probability
in [0,1] to trust value in [–1,1], using
information theoretic approach.
Y. L. Sun, et al, 2/2008

Information Theoretic Interpretation
of Trustworthiness Probability

• T(trustee : trustor, action) =
if 0.5 <= p
then 1 – H(p) /* 0.5 <= p <= 1 */
else H(p) – 1 /* 0 <= p <= 0.5 */
where
H(p) = – p log2(p) – (1 – p) log2(1 – p)


Plot of T(trustee : trustor, action) vs. p

Trust portion (p in [0.5,1]) Distrust portion (p in [0,0.5])


Linear vs Nonlinear Map
Relative to computing trust as
T(trustee : trustor, action) = ( p – 0.5 ) * 2
to map of trust probability p in [0,1] to a
trust value in [-1,+1], the information
theoretic formulation yields a non-linear
map that amplifies the effect of changes
to trust probability on the trust value at
the extremes.

Direct Trust : Functional
Policy-based Process
(Using Trustworthiness Qualities)


General Approach to Trust Assessment
• Domain dependent qualities for determining
trustworthiness
– Based on Content / Data
– Based on External Cues / Metadata

• Domain independent mapping to trust values
or levels
– Quantification through abstraction and
classification

Example: Wikipedia Articles
• Quality (content-based)
– Appraisal of information provenance
• References to peer-reviewed publication
• Proportion of paragraphs with citation
– Article size
• Credibility (metadata-based)
– Author connectivity
– Edit pattern and development history
• Revision count
• Proportion of reverted edits - (i) normal (ii) due to vandalism
• Mean time between edits
• Mean edit length.
Sai Moturu, 8/2009

(cont’d)
• Quantification of Trustworthiness
– Based on Dispersion Degree Score
(Extent of deviation from mean)
• Evaluation Metric
– Ranking based on trust level (determined from
trustworthiness scores), and compared to gold
standard classification using Normalized
Discounted Cumulative Gain (NDCG)
• RATINGS: featured, good, standard, cleanup, and stub.
• NDCG: penalizes more heavily errors at the top.


Example: Websites
• Trustworthiness estimated based on criticality
of data exchanged.
• Email address / Username / password
• Phone number / Home address
• Date of birth
• Social Security Number / Bank Account Number
• Intuition: A piece of data is critical if and only
if it is exchanged with a small number of
highly trusted sites.

Indirect Trust : Referral + Functional
Variety of Trust Metrics and Models
(Using Propagation – Chaining and Aggregation over Paths)


Collaborative Filtering
• Collaborative Filtering: Item-rating by a user
predicted on the basis of user’s similarity to
other users.

• Similarity Measures:
• Profile-based
• Item-ratings-based
• Item-category-based


Collaborative Filtering
• Pros:
– Items-agnostic
– Scales well over time with large number of items
• Cons:
– Data Sparsity Problem: Small number of
common items between users.
– Cold Start Users: Small number of items rated
by a user.
– Prone to Copy-Profile Attack: An attacker can
create a targeted-user-like profile to
manipulate recommendations.
Massa-Avesani, 2007

Trust-aware Recommender System
• TaRS uses explicit/direct trust between users to
predict implicit/indirect trust between users
through chaining.
• Collaborative Filtering Limitations Overcome:
– Mitigates Data Sparsity: Trust propagation is more
general and improves coverage.
– Bootstraps Cold Start Users: A single trust link from a
new user can enable the user to inherit several
“parental” recommendations.
– Robust w.r.t Copy-Profile Attack: Fake identities are
not trusted by an active user.
Massa-Avesani, 2007

Trust Propagation Frameworks
• Chaining, Aggregation, and Overriding
Golbeck – Hendler, 2006 Massa-Avesani, 2005 Sun et al, 2006
Bintzios et al, 2006 Thirunarayan et al, 2009
• Trust Management Richardson et al, 2003
• Abstract properties of operators
• Reasoning with trust Guha et al., 2004
• Matrix-based trust propagation
• The Beta-Reputation System
• Algebra on opinion = (belief, disbelief, uncertainty)
Josang and Ismail, 2002

Trust Propagation Algorithms
• Top-down
• 1: Extract trust DAG (eliminate cycles)
• 2: Predict trust score for a source in a target by
aggregating trust scores in target inherited
from source’s “trusted” parents weighted with
trust value in the corresponding parent.
–Computation is level-by-level
–Alternatively, computation can be based on
paths.
Golbeck – Hendler, 2006


Trust Propagation Algorithms
• Bottom-up
• 1: Extract trust DAG (eliminate cycles)
• 2: Predict trust score for a source in a target by
aggregating trust scores in target inherited
from target’s “trusted” neighbors weighted
with trust value in the corresponding neighbor.
–Computation is level-by-level
–Alternatively, computation can be based on
paths. Massa-Avesani, 2005
Bintzios et al, 2006


Top-down vs Bottom-up (visualized)

Target

w1 w3
Target w2

Trust-Network

w1 w2 w3

source source


Example: Comparative Analysis
Different Interpretation:
q distrusts s (Bintzios et al’s)
vs
q has no information about
the trustworthiness of s (our’s,
Golbeck rounding algorithm)


Example: Well-founded
Cyclic Trust Network

Thirunarayan and Verma, 2007


Example: Using TidalTrust Algorithm


Indirect Trust : Referral + Functional
Variety of Bayesian Trust Models
With Applications to Mobile Ad hoc Networks
Wireless Sensor Networks, etc.


• Trust link: {subject : agent, action}
• For MANETs (resp. cognitive radio)
– Functional => Packet Forwarding (resp. quality of
spectrum / channel)
– Referral => Recommendations
Based on beta-reputation model:
Probability for trust = (S + 1) / (S + F +2)
where S = Number of good actions
F = Number of bad actions


• Direct Trust for Packet Forwarding
• S = Number of packets forwarded
• F = Number of packets dropped
• S + F = Total number of requests for packet forwarding
• Direct Trust for Recommendations
• S = Number of times observed direct trust for packet
forwarding approximates expected indirect trust for
packet forwarding (trust over transit path : r+f)
• F = Number of times observed direct trust for packet
forwarding does not approximate expected indirect
trust for packet forwarding (trust over transit path : r+f)

Indirect Trust : Functional and Referral
• Indirect Trust for Packet Forwarding
– Used when direct trust is not available
»(overriding behavior)
• Chain links for a path from a recommender to the target
– Multiplicative
• Aggregate over multiple (parallel) paths from
recommenders to the target
– Unclear, in general

• Indirect Trust for Recommendations
• Obtained implicitly through computed referral trust


Trust Propagation Rules :
Axioms for Trust Models

Rule 1: Concatenation
propagation does
not increase trust. |T(A1,C1)| <= min(|R(A1,B1)|, |T(B1,C1)|)

Rule 2: Multipath
propagation does
not reduce trust. 0<=T(A1,C1) <= T(A2,C2) for R1 > 0 and T2 >= 0
0>=T(A1,C1) >= T(A2,C2) for R1 > 0 and T2 < 0

May 21, 2012 Trust Networks: T. K. Prasad
Sun et al, 2006 93

(cont’d)
Rule 3: Trust based on multiple referrals from a
single source should not be higher than that
from independent sources.

0<=T(A1,C1) <= T(A2,C2) for R1, R2, R3 > 0 and T2 >= 0
0>=T(A1,C1) >= T(A2,C2) for R1, R2, R3 > 0 and T2 < 0

Sun et al, 2006 94

Implementation
1 2

Sun et al, 2006 95

Trust Paths Visualized for Scalability:
Semantics unclear based on Sun et al’s spec

Bottom-up computation
reflects our needs better?


Trust : Functional and Referral
• Direct Trust for Primitive Actions based on
• S = Number of success actions
• F = Number of failed actions
• S + F = Total number of actions
• Indirect Trust via Recommendations based on
summing direct experiences of recommenders
• Sk = Number of success actions for kth recommender
• Fk = Number of failed actions for kth recommender

• No chaining for referrals Denko-Sun 2008


Cumulative Trust using
Direct Experience and Recommendations
• Cumulative Trust is obtained by using total
number of success actions and failed actions
from direct experience (ns,nu) and from i
(indirect experiences through)
recommendations (nsr,nur).


Contents of [Ganeriwal et al, 2007] Paper

• (a,b)-parameters to compute trust of i in j is
obtained by combining direct observations
(aj,bj) with indirect observations (ajk,bjk) from
k weighted by (ak,bk) using [Josang-Ismail,
2002] chaining/discounting rule.
• Obtains cumulative trust by combining direct trust from
a functional link and indirect trusts using paths
containing one referral link and one functional link.
• However, it does not distinguish functional and referral
trust.


Beta Reputation System
1 2

107

Beta Reputation System
1 2

Rule 1: Concatenation propagation (reputation
discounting)
r = 2*r1 * r2 / (s1+2)(r2+s2+2)+2*r1
s = 2*r1 * s2 / (s1+2)(r2+s2+2)+2*r1
Rule 2: Multipath propagation (combining
feedback)
r = r1 + r2 s = s1 + s2
Rule *: Temporal Decay (Forgetting)
108

Contents of [B-Trust, 2006] Paper
• Uses generic K-level discrete trust metric (as
opposed to 2-level metric)
– E.g., (very untrustworthy, untrustworthy, trustworthy, very trustworthy)
– E.g., reminiscent of Amazon recommendation ratings
• Distributed (local), robust, lightweight,
computational trust that takes into account
context, subjectivity, and time
– a la reputation-based approach
• Application: Pervasive computing

BUG -> FLAWED LEARNING?

• The approach does not clearly separate the
use of stable background knowledge for
applying Bayes’ rule, from the need to
dynamically learn background knowledge for
gleaning trust from experience.
• As a result, the initial trust values do not
change in response to experience.


Comparative Analysis
[Menko-T.Sun] : Beta-distribution based
• Direct functional trust and indirect functional
trust (through direct referrals).
– Trivial chaining.
– One fixed context, local and distributed.
• Robustness improved by dropping extreme
recommendations, though recommenders not
distinguished.


[Ganeriwal et al]: Beta-distribution based
• Functional and Referral trust mixed up.
– Context glossed over, local and distributed.
• Robustness improved by chaining trust links of
length 2, using Josang-Ismail opinion
composition.
– Recommenders distinguished.
– Chaining weighs recommendations by
recommender trust.

[Y. Sun et al]: Beta-distribution based
• Functional and Referral trust separated.
• One Context, hybrid (dynamically formed trust
network) and distributed.
• Information-theoretic approach
• Axiomatic specification and implementation of
trust propagation (chaining and aggregation)
• Incomplete w.r.t. arbitrary trust networks
• Robustness and Quality improved by analyzing
dynamically formed trust network.

[B-Trust et al]: Multi-valued trust - Bayesian
– Context-based, local and distributed.
• Individualized aggregation, trivial chaining.
• Nice roadmap for theory, specification, and
implementation of trust networks
– Multi-valued Trust evolution : novel but buggy.


[MLT-Approach]: Multi-level trust using Dirichlet
Distribution
– Context-based, local and distributed.
• Individualized aggregation, trivial chaining.
• Based on B-Trust roadmap but MLT evolution
based on Dirichlet distribution: conceptually
satisfactory and computationally efficient
• Example-based analysis for insights

Security Issues:
Threats and Vulnerabilities
Attacks and Robustness Analysis


Attacks
• Trust Management is an attractive target for
malicious nodes.
 Bad mouthing attack (Defamation)
 Dishonest recommendations on good nodes (calling them
bad)
 Ballot stuffing attack (Collusion)
 Dishonest recommendations on bad nodes (calling them
good)
 Sybil attack
 Creating Fake Ids
 Newcomer attack
 Registering as new nodes


Attacks
• Inconsistency in time-domain
 On-Off attack
 Malicious node behaves good and bad alternatively to
avoid detection
 Sleeper attack
 Malicious node acquires high trust by behaving good
and then strikes by behaving bad
 Inconsistency in node-domain
 Conflicting Behavior Attack
 Provide one recommendation to one set of peers and a
conflicting recommendation to a disjoint set of peers


Security : Robustness w.r.t Attacks
 Bad mouthing attack
 Example: Competent nodes downplay competitions.
 Example: Can diminish throughput due to lost capacity.
 Approach:
 Separate functional and referral trust, updating
referral trust to track good recommendations
 Trust composition rules ensure that low or
negative referral trust does not impact decision
 Low trust nodes can be branded as malicious and
avoided. (Not viable if majority collude.)


 Ballot stuffing attack
 Example: Malicious nodes collude to recommend each
other.
 Example: Can cause unexpected loss of throughput.
 Approach:
 Feedback : Cross-check actual functional
performance with expected behavior via referral,
and update (reward/penalize) referral trust (in
parent) accordingly (in addition to updating
functional trust (in target))

Security : Robustness w.r.t. Attacks
 Sybil attack
 Create Fake Ids to take blame for malicious
behavior (dropping packets)
 Newcomer attack
 Register as new node to erase past history
 Approach
 Requires separate (key-based or security token-
based) authentication mechanism (with TTP) to
overcome these attacks.


 On-Off attack
 Sleeper attack
 Example: Due to malice or environmental changes
 Approach:
 Use forgetting factor (0<=b<=1):
k good/bad actions at t1
= k * b(t2 – t1) good/bad actions at t2 (> t1)


Forgetting Factor
k good/bad actions at t1 = k * b(t2 – t1) good/bad actions at t2 (> t1)

• High b value (0.9) enhances memorized time
window, while low b value (0.001) reduces it.
– High b enables malicious nodes (on-off/sleeper
attackers) to use prior good actions to mask
subsequent intentional bad actions.
• Reduces reliability.
– Low b forces legitimate nodes to be avoided due
to short spurts of unintentional bad actions.
• Reduces throughput.


Adaptive Forgetting Factor
• Intuition: Bad actions are remembered for a
longer duration than good actions.
• Actions performed with high trust forgotten quicker than
actions performed with low trust.
Choose b equal to ( 1 – p )
Choose b = 0.01 when p in [0.5,1] else 0.9
 Example: Similar ideas used in Ushahidi
 Note: Effectively, more good actions are
necessary to compensate for fewer bad actions,
to recover trust.


 Malicious node divide and conquer, by behaving
differently (resp. by providing different
recommendations) to different peers, causing
peers to provide conflicting recommendations to
source about the malicious node (resp. about
some target), reducing source’s referral trust in
some peers.
 Eventually, this causes recommendations of some peers
to be ignored incorrectly.


Example
• Peer Node Set 1: 1, 2, 3, and 4
• Peer Node Set 2: 5, 6, 7, and 8
• Malicious node 0 behaves well towards nodes in
Set 1 but behaves badly towards nodes in Set 2.
• When node 9 seeks recommendations from
nodes in Set 1 U Set 2 on node 0, node 9 receives
conflicting recommendations on malicious node
0, causing referral trust in nodes in Set 1 or
nodes in Set 2 to be lowered.
=> Eventually throughput lowered



 Issue: Can recommenders get feedback
to reduce trust in malicious node?
Otherwise, referral trust cannot be
relied upon for detecting malicious
nodes.


 If cumulative referral trust in B is computed using
direct experiences of several recommenders,

then it is possible to weed out extreme
experiences using deviation from the mean trust
value, where S is some chosen threshold.

Denko-Sun 2008 128

Contents of [Ganeriwal et al, 2007] Paper

• Combined trust of i in j is obtained from direct
observations (aj,bj) of j by i and indirect
observations (ajk,bjk) from k to i weighted by
(ak,bk) using [Josang-Ismail, 2002]
chaining/weighting/discounting rule.
• This discounting rule makes the local trust
computation resilient to bad mouthing ajk << bjk
and ballot stuffing ajk >> bjk attacks from
unreliable/malicious nodes ak << bk.
• It requires aging to be resilient to sleeper attacks.

APPROACH/ Trust Type / Trust Model / Robustness to
METRIC Context Foundation Attacks
D[3] / Functional / One Trivial chaining / Ballot-stuffing;
Binary Beta-PDF Bad-mouthing

G[4] / Functional / Josang-Ismail Ballot-stuffing;
Binary Indistinguishable discounting / Bad-mouthing;
Beta-PDF Sleeper and On-
off
S[6] / Functional + Referral Limited chaining Ballot-stuffing;
Binary / One and aggregation / Bad-mouthing;
Beta-PDF Sleeper and On-
off
Q[28] / Multi-level Functional + Referral / No / Ballot-stuffing;
Multiple Bayesian Bad-mouthing;
Sleeper and On-
Ad Hoc
off; Sybil

Ours / Functional + Referral / No / Ballot-stuffing;
Multi-level Multiple Dirichlet-PDF Bad-mouthing;
Sleeper and On-
off; Conflicting
behavior


Research Challenges

(What-Why-How of trust?)

HARD PROBLEMS


Generic Directions

• Finding online substitutes for traditional cues
to derive measures of trust.

• Creating efficient and secure systems for
managing and deriving trust, in order to
support decision making.
Josang et al, 2007


Robustness Issue
You can fool some of the
people all of the time, and all
of the people some of the
time, but you cannot fool all
of the people all of the time.
Abraham Lincoln,
16th president of US (1809 - 1865)

Trust : Social Networks vs
Machine Networks
• In social networks such as Facebook, trust is
often subjective, while in machine networks
and social networks such as Twitter, trust can
be given an objective basis and approximated
by trustworthiness.
• Reputation is the perception that an agent
creates through past actions about its
intentions and norms.
– Reputation can be a basis for trust.


Sensor Networks


Our Research
Abstract trustworthiness of sensors and
observations to perceptions to obtain actionable
situation awareness!
Web
T

observe perceive

T T

“real-world”


Trust

Strengthened Trust


Concrete Application
• Applied Beta-pdf to Mesowest Weather Data

– Used quality flags (OK, CAUTION, SUSPECT)
associated with observations from a sensor
station over time to derive reputation of a sensor
and trustworthiness of a perceptual theory that
explains the observation.

– Perception cycle used data from ~800 stations,
collected for a blizzard during 4/1-6/03.

Concrete Application
• Perception Cycle
– http://harp.cs.wright.edu/perception/

• Trusted Perception Cycle
– http://www.youtube.com/watch?v=lTxzghCjGgU


Mean of beta pdf vs. Time (for stnID = SBE)
1.2

1

0.8

Mean Beta
0.6
Value

0.4

0.2

0
3/31/2003 0:00 4/1/2003 0:00 4/2/2003 0:00 4/3/2003 0:00 4/4/2003 0:00 4/5/2003 0:00 4/6/2003 0:00 4/7/2003 0:00
Time


Research Issues
• Outlier Detection
– Homogeneous Networks
• Statistical Techniques
– Heterogeneous Networks (sensor + social)
• Domain Models
• Distinguishing between abnormal phenomenon
(observation), malfunction (of a sensor), and
compromised behavior (of a sensor)
– Abnormal situations
– Faulty behaviors
– Malicious attacks Ganeriwal et al, 2008


Social Networks


Our Research

• Study semantic issues relevant to trust

• Proposed model of trust/trust metrics to
formalize indirect trust


Quote
• Guha et al:
While continuous-valued trusts are
mathematically clean, from the standpoint
of usability, most real-world systems will
in fact use discrete values at which one
user can rate another.
• E.g., Epinions, Ebay, Amazon, Facebook, etc all
use small sets for (dis)trust/rating values.


Our Approach

 Trust formalized in terms of partial orders
(with emphasis on relative magnitude)
 Local but realistic semantics
 Distinguishes functional and referral trust
 Distinguishes direct and inferred trust

 Direct trust overrides conflicting inferred trust

 Represents ambiguity explicitly

Thirunarayan et al , 2009


Formalizing the Framework

• Given a trust network (Nodes AN, Edges RL U
PFL U NFL with Trust Scopes TSF, Local
Orderings ⪯ANxAN), specify when a source can
trust, distrust, or be ambiguous about a
target, reflecting local semantics of:
• Functional and referral trust links
• Direct and inferred trust
• Locality

(In recommendations)

(For capacity to act)

(For lack of
capacity to act)

Similarly for Evidence in support of Negative Functional Trust.


Benefits of Formal Analysis
• Enables detecting and avoiding
unintended consequences.
– An earlier formalization gave priority to “certain“
conclusion from less trustworthy source over
“ambiguous“ conclusion from more trustworthy
source.
The whole problem with the world is
that fools and fanatics are always so
certain of themselves, but wiser people
so full of doubts. — Betrand Russell

Practical Issues
• Refinement of numeric ratings using
reviews in product rating networks
– Relevance : Separate ratings of vendor or about
extraneous features from ratings of product
• E.g., Issues about Amazon’s policies
• E.g., Publishing under multiple titles (Paul Davies’ “The Goldilock’s
Enigma” vs. “Cosmic Jackpot”)
– Polarity/Degree of support: Check consistency
between rating and review using sentiment
analysis; amplify hidden sentiments
• E.g., rate a phone as 1-star because it is the best 


Research Issues
• Determination of trust / influence from
social networks
– Text analytics on communication
– Analysis of network topology
• E.g., follower relationship, friend relationship, etc.
• Determination of untrustworthy and
anti-social elements in social networks
• HOLY GRAIL: Direct Semantics in favor of
Indirect Translations

Research Issues
• Improving Security : Robustness to Attack
– How to exploit different trust processes to detect
and recover from attacks?
• Bad mouthing attack
• Ballot stuffing attack
• Sleeper attack
– Temporal trust discounting proportional to trust value
– Using policy-based process to ward-off attack using
reputation-based process
• Sybil attack
• Newcomer attack

Research Issues
• Intelligent integration of mobile sensor and
social data for situational awareness
– To exploit complementary and corroborative
evidence provided by them
– To obtain qualitative and quantitative context
– To improve robustness and completeness
– To incorporate socio-cultural, linguistic and
behavioral knowledge as part of ontologies to
improve semantic processing and analysis of data


Complementary and Corroborative
Information
Sensors observe
slow moving
traffic

Complementary
information
from social
networks


Corroborative Evidence

Evidence for
reported
observations


Interpersonal and Ecommerce
Networks


Research Issues
• Linguistic clues that betray
trustworthiness
• Experiments for gauging interpersonal
trust in real world situations
– *Techniques and tools to detect and amplify
useful signals in Self to more accurately predict
trust and trustworthiness in Others

*IARPA-TRUST program


Research Issues
• Other clues for gleaning trustworthiness
– Face (in photo) can effect perceived
trustworthiness and decision making
–Trust-inducing features of e-commerce
sites can impact buyers
– Personal traits: religious beliefs, age,
gullibility, benevolence, etc
– Nature of dyadic relationship

Research Issues
• Study of cross-cultural differences in
trustworthiness qualities and trust thresholds
to better understand
– Influence
• What aspects improve influence?
– Manipulation
• What aspects flag manipulation?


Collaborative Systems :
Grid and P2P Computing


Research Issues
• Trust-aware resource management and
scheduling
– Clients specify resource
preferences/requirements/constraints
• Trust models for P2P systems
– To detect bad domains
– To detect bogus recommendations and attacks
Azzedin and Maheshwaran, 2002-2003
Azzedin and Ridha, 2010
Bessis et al, 2011


Bayesian Trust Management Framework :
Multi-level Trust Metric

Illustrating a General Approach

Quercia et al 2006
Josang and Haller 2007
May 21, 2012 Trust Networks: T. K. Prasad Thirunarayan et al 2012 170

Outline
• Motivation : Multi-level trust management
• Mathematical Foundation: Dirichlet
Distribution
• Implementation and Behavior Details:
– Local Trust Data Structures
– Trust Formation
– Bayesian Trust Evolution
• Analysis of Robustness to Attacks: Security
• Evaluation: Example trace vs. experiment

Motivation

• Uses K-level discrete trust metric

– E.g., Amazon’s 5-star trust metric can be
interpreted as signifying (very untrustworthy,
untrustworthy, neutral, trustworthy, very
trustworthy) or (very dissatisfied, dissatisfied,
neutral, satisfied, very satisfied).


Approach
• Multi-level trust management approach
formalizes a distributed, robust,
lightweight, computational trust that
takes into account context, subjectivity,
and time.
• Applies Dirichlet distribution, a
generalization of Beta-distribution.


Dirichlet Distribution


K-level Trust Metric
• K-level trust probability vector:
x = (x1, . . ., xK)
where (x1 + . . . + xK = 1).

• Example: If a 5-star rating system has 50 people
giving 5-stars, 20 people giving 4-stars, 5 people
giving 3-stars, 5 people giving 2-stars, and 20 people
giving 1-star, then the 5-level trust metric probability
vector is (0.5,0.2,0.05,0.05,0.2).


Trust and Experience
• Experience is a realization of latent trust and
helps predicting trust.
• Probability of an experience-level sequence,
with a1 - 1 counts of level 1 experience, …, aK - 1
counts of level K experience is:

* ( (a1 +…+ aK – K) ! / (a1-1 ! *…* aK-1 !) )


Dirichlet Distribution
• The Dirichlet distribution is the probability
density function for x = (x1, . . ., xK) given
(a1,…,aK):


Why use Dirichlet Distribution?
• If the prior distribution of x is uniform, then
the Dirichlet family of distribution shown
below gives posterior distribution of x after
ai-1 occurrences of level i experience with
probability xi, for each i in [1, K]:


• Dirichlet distribution is a conjugate prior for
multinomial distribution.
• Consequence:
– Estimated distribution updated for a new experience
at level i, by just incrementing ai parameter.
– In contrast: if prior distribution is different from
Dirichlet, then it is conceptually hard to comprehend
and computationally inefficient to compute posterior
distribution, in general.
– Icing on the cake: Uniform distribution (signifying
ignorance) is Dirichlet!

Dirichlet distribution is a conjugate prior
for multinomial distribution.


• Convenient Abstraction
–Abstraction of K-level Dirichlet
distribution by combining different
levels still yields Dirichlet distribution
with the corresponding parameters
merged.
• Conceptually and computationally
pleasing property


Visualizing Dirichlet Distribution (K=3):
Color Density plot on 2D simplex


Dynamic Trustworthiness
• Best estimate of trust for Dir(a1,...,aK) (gleaned
from (ai-1) experiences at level i, for all i in [1,K])
is the mean vector (a1/a0,…,aK/a0), and the
associated confidence is the variance vector.


Implementation and Behavior Details


Local Data Structures

• To store relevant information to compute
direct (functional) and indirect (referral) trust.

• Each node maintains locally, for each peer and
each context, four vectors of length K.



• Direct Trust Vector: Peers X Contexts X Peers -> Probability-Vector-K

• dtv(px,c,py) = (d1,d2,…,dK)

• Direct Experience Matrix: Peers X Contexts X Peers -> Count-Vector-K

• dem(px,c,py) = (ec1,…,ecK)



• Recommended Trust Vector: Peers X Contexts X Peers -> Probability-
Vector-K

• rtv(px,c,py) = (r1,r2,…,rK)
• Sent Recommendation Matrix: Peers X Contexts X Peers -> Count-Vector-K

• srm(px,c,py) = (sr1,…,srK)



• Initialization: To reflect complete
ignorance via uniform distribution, we
set the probability vectors dtv and rtv to
(1/K,…,1/K), and the elements of the
count vector dem and srm to (0,…,0).
• These are Dirichlet distributed in the
limiting case where ai’s are 1.

Trust Formation
• Overall trust vector is weighted combination
of direct trust vector and recommended trust
vector.
• Weights determined using
– Objective confidence values using variance
(deviation from the mean)
– Subjective relative preference for direct
experience over recommendations
• Dependence on recommended trust yet to be explored


Trust Decision

• Assuming that trust-level scale is linear, the
trust distribution vector (d1,d2,…,dK) can be
mapped to the closed interval [0,1], or to
consolidated trust level, in order to act.

• Trust threshold should be determined based
on the context and risk tolerance / disposition
/ propensity to trust.

Trust Evolution

• Direct (recommended) trust vectors are updated
for a new experience (recommendation).
• Key Idea: Dirichlet distribution is the conjugate
prior of the multinomial distribution. So it is
adequate to maintain counts of direct experience
and sent recommendations, to best estimate
direct trust and recommended trust vectors
respectively.

Trust Evolution
• Simple Scheme (Direct Trust)
For a new experience at level i,
dem(px,x,py) = (ec1,…,ecK) becomes
demnew(px,x,py) = (ec1,…, eci+1,…,ecK)
and dtv(px,c,py) becomes
dtvnew(px,c,py) = (d1,d2,…,dK)
where di = eci+1 / (ec1 + … + eck+1) and
dj = ecj / (ec1 + … + eck+1)
for each j in [1,K] and j =/= i.


Trust Evolution
• Robust Scheme
To incorporate differential aging of experience
counts as a function of their level (and to
incorporate “long term memory for low-level
experience and short term memory for high-
level experience”), we use a decay vector
(l1,…,lK), where 1 >= l1 >= … >= lK > 0, that
modifies update rule as:


Trust Evolution
• Robust Scheme (Direct Trust)
For a new experience at level i,
demnew(px,x,py) = (ec1,…, eci + 1,…,ecK).
For every clock tick (with context-based delay),
demnew(px,x,py) = (l1*ec1,…, lK*ecK)


Trust Evolution
• Robust Scheme (Direct Trust)
For every clock unit and new experience,
dtv(px,c,py) becomes
dtvnew(px,c,py) = (d1,d2,…,dK)
where di = eci / (ec1 + … + eck)
for each i in [1,K].
• Subtlety: Experience counts should saturate at
1 rather than diminish to 0 with time. (See code)

Trust Evolution Illustrated

Final Trust Distribution Final Trust Distribution
Experience Sequence
(Simple Scheme) (Robust Scheme)

[1,1,1] (0.57,0.14,0.14,0.14)
(0.55,0.15,0.15,0.15)

[1,4,1,4] (0.375,0.125,0.125,0.375)
(0.42,0.14,0.14,0.29)

[1,1,4,4,4,4,1,1] (0.42,0.08,0.08,0.42)
(0.5,0.1,0.1,0.3)

[1,1,4,4,4,4,1,1,1] (0.53,0.07,0.07,0.33)
(0.64,0.1,0.1,0.17)

[2,3,2,3] (0.125,0.375,0.375,0.125)
(0.16,0.4,0.3,0.14)


Trust Evolution Illustrated
Experience Trust Distribution Trace Trust Distribution Trace Beta-pdf
Sequence Value (Simple Scheme) (Robust Scheme) (cf. n =2)
(0.25,0.25,0.25,0.25) (0.25,0.25,0.25,0.25) 0.5

1 (0.4,0.2,0.2,0.2) (0.4,0.2,0.2,0.2) 0.33

1 (0.5,0.17,0.17,0.17) (0.53,0.165,0.155,0.15) 0.25

1 (0.57,0.14,0.14,0.14) (0.55,0.15,0.15,0.15) 0.2

n (0.5,0.125,0.125,0.25) (0.5,0.13,0.12,0.25) 0.33

n (0.44,0.11,0.11,0.33) (0.46,0.135,0.135,0.27) 0.43

n (0.4,0.1,0.1,0.4) (0.42,0.12,0.11,0.35) 0.5

n (0.36,0.1,0.1,0.45) (0.37,0.12,0.12,0.38) 0.55

1 (0.42,0.08,0.08,0.41) (0.47,0.11,0.11,0.31) 0.5

1 (0.46,0.08,0.08,0.38) (0.53,0.11,0.11,0.24) 0.45

1 (0.5,0.07,0.07,0.35) (0.6,0.1,0.1,0.2) 0.41

1 (0.53,0.07,0.07,0.33) (0.65,0.1,0.1,0.14) 0.38

n (0.5,0.0625,0.0625,0.375) (0.6,0.1,0.1,0.2) 0.43

1 (0.53,0.06,0.06,0.35) (0.64,0.1,0.1,0.17) 0.4


Analysis and Robustness Issues


Salient Properties
• Symmetry
– Simple Scheme is symmetric w.r.t.
trust/experience levels while Robust Scheme is
somewhat asymmetric because of non-uniform
decay.
– Experience levels are “preserved” in that
extreme/controvertial behavior (credulous
interpretation) is treated differently from
ignorance (skeptical interpretation).


Salient Properties
• Effect of Order of Experience

– Simple Scheme is sensitive to the counts of
various experience levels, but not to the order of
experience.

– Robust Scheme is sensitive to the order of
experience.


Salient Properties
• Differential Aging of experience levels
• It exhibits limited and selective
memory.
–It retains low-level experiences
much longer than high-level
experiences.
»Parameters: Decay rate and
saturation

Related Work on Multi-level Trust
with Applications
The described approach is similar to Dirichlet Reputation
System [Josang-Haller, 2007].
Applications:
• Browser toolbar for clients to see the user ratings and
for users to provide ratings (critical surfer model)
[Josang-Haller, 2007]
• Evaluating partners in Collaborative Environments
[Yang and Cemerlic, 2009]
• Formalizing Multi-Dimensional Contracts [Reece, et al,
2007]
• In Collaborative Intrusion Detection System [Fung et al,
2011 ]


Conclusion
• Provided simple examples of trust (Why?)
• Explained salient features of trust (What?)
• Showed examples of gleaning trustworthiness
(How?)
• Touched upon research challenges in the
context of
• Sensor Networks
• Social Networks
• Interpersonal Networks
• Collaborative Environments

Holy Grail
for Automatic Trust Computation

Develop expressive trust
networks that can be assigned
objective semantics.


Thank You!


Trust networks

More Related Content

Similar to Trust networks

Recently uploaded

Trust networks