Trust in Pervasive Social Computing Licia Capra Dept of Computer Science University College London

EVOLUTION OF MOBILE TECHNOLOGY

TRANSFORMATION OF INTERNET USERS

A DIGITAL TAPESTRY

PERVASIVE SOCIAL COMPUTING

A DIGITAL TAPESTRY Millions of different resources available to consumers Content overload

Millions of different resources available to consumers

Content overload

A DIGITAL TAPESTRY Users need help to find interesting content An enormous market for niche content is appearing

Users need help to find interesting content

An enormous market for niche content is appearing

CONNECTING SYPPLY & DEMAND Filters are essential to connect supply and demand

Filters

are essential

to connect

supply

and

demand

COLLABORATIVE FILTERING Quantify user similarity Similar users are deemed competent Limitations Create recommendations based on behaviour of similar users Mass market Subject to abuse

Quantify user similarity

Similar users are deemed competent

Limitations

Create recommendations based on behaviour of similar users

Mass market

Subject to abuse

SOCIAL NETWORKS Limitations No subjectivity No transitivity Determine user friendship Friends are deemed to have good intent Share content across friendship relations

Limitations

No subjectivity

No transitivity

Determine user friendship

Friends are deemed to have good intent

Share content across friendship relations

BROWSING THE TAPESTRY Joint work with Matteo Dell’Amico @ Eurecom Presented at the Joint iTrust and PST Conferences on Privacy, Trust Management and Security, June 08

BROWSING THE TAPESTRY – OUTLINE Philosophy of the Approach Intent Competence Social Filtering Realisation of the Approach Intent: Personalised PageRank Competence: HITS Social Filtering: SOFIA Experiments Accuracy Robustness

Philosophy of the Approach

Intent

Competence

Social Filtering

Realisation of the Approach

Intent: Personalised PageRank

Competence: HITS

Social Filtering: SOFIA

Experiments

Accuracy

Robustness

BROWSING THE TAPESTRY – OUTLINE Philosophy of the Approach Intent Competence Social Filtering Realisation of the Approach Intent: Personalised PageRank Competence: HITS Social Filtering: SOFIA Experiments Accuracy Robustness

Philosophy of the Approach

Intent

Competence

Social Filtering

Realisation of the Approach

Intent: Personalised PageRank

Competence: HITS

Social Filtering: SOFIA

Experiments

Accuracy

Robustness

PHILOSOPHY OF THE APPROACH Intent

Intent

PHILOSOPHY OF THE APPROACH Intent: willingness to provide honest judgements A B C D Direct Trust Inferred Trust Web of Trust: social network where A is connected to B if A trusts that B behaves honestly Transitivity Pattern: propagate trust over intent (e.g., “I trust the friends of my friends”) Limitation: transitivity does not cater for the taste of users (i.e., subjectivity in niche markets)

Intent: willingness to provide honest judgements

Web of Trust: social network where A is connected to B if A trusts that B behaves honestly

Transitivity Pattern: propagate trust over intent (e.g., “I trust the friends of my friends”)

Limitation: transitivity does not cater for the taste of users (i.e., subjectivity in niche markets)

PHILOSOPHY OF THE APPROACH Competence

Competence

PHILOSOPHY OF THE APPROACH Competence: ability to provide correct judgements A B X Y Direct Trust Inferred Trust Bipartite graph: network of judgements where user A is connected to judgement X if A expressed X Co-citation Pattern: propagate trust over competence (e.g., “I agree with judgements of competent users”) Limitation: co-citation subject to abuse (i.e., Sybil attack, profile injection, shilling, etc.)

Competence: ability to provide correct judgements

Bipartite graph: network of judgements where user A is connected to judgement X if A expressed X

Co-citation Pattern: propagate trust over competence (e.g., “I agree with judgements of competent users”)

Limitation: co-citation subject to abuse (i.e., Sybil attack, profile injection, shilling, etc.)

= Intent & Competence

PHILOSOPHY OF THE APPROACH Social Filtering Pattern A can infer trust for judgement Y expressed by D if There exists a directed path from A to D in the web of trust D is trusted to be well intentioned (robustness) A and expressed at least one common judgement D is trusted to be competent (accuracy) A D X Y Direct Trust for Judgements Direct Trust for Users C B Inferred Trust for Judgements

Social Filtering Pattern

A can infer trust for judgement Y expressed by D if

There exists a directed path from A to D in the web of trust

D is trusted to be well intentioned (robustness)

A and expressed at least one common judgement

D is trusted to be competent (accuracy)

BROWSING THE TAPESTRY – OUTLINE Philosophy of the Approach Intent Competence Social Filtering Realisation of the Approach Intent: Personalised PageRank Competence: HITS Social Filtering: SOFIA Experiments Accuracy Robustness

Philosophy of the Approach

Intent

Competence

Social Filtering

Realisation of the Approach

Intent: Personalised PageRank

Competence: HITS

Social Filtering: SOFIA

Experiments

Accuracy

Robustness

REALISATION OF THE APPROACH Evaluating Intent Desirable properties Longer paths disperse trust Adding paths increase trust Limit amount of trust propagated through attack edges X A B C D PageRank Goal: rank importance of web pages Intuition: an authoritative page is linked by many authoritative pages Swap WWW graph with Web of Trust Goal: rank trustworthiness of users Intuition: an honest user is being trusted by many honest users Lawrence Page, Sergey Brin, Rajeev Motwani, Terry Winograd. “ The PageRank Citation Ranking: Bringing Order to the Web”. Stanford Digital Library. 1998

Evaluating Intent

Desirable properties

Longer paths disperse trust

Adding paths increase trust

Limit amount of trust propagated through attack edges

PageRank

Goal: rank importance of web pages

Intuition: an authoritative page is linked by many authoritative pages

Swap WWW graph with Web of Trust

Goal: rank trustworthiness of users

Intuition: an honest user is being trusted by many honest users

REALISATION OF THE APPROACH Evaluating Intent Personalised PageRank Start random walk at (honest) evaluating node A Jump from X to Y with probability proportional to edge weight Stop with probability 1-  Importance of  Max rank for Sybil region: p /(1-  ) Low  implies shorter paths, higher robustness, faster convergence but we don’t trust honest users when they are socially far away

Evaluating Intent

Personalised PageRank

Start random walk at (honest) evaluating node A

Jump from X to Y with probability proportional to edge weight

Stop with probability 1- 

Importance of 

Max rank for Sybil region: p /(1-  )

Low  implies shorter paths, higher robustness, faster convergence but we don’t trust honest users when they are socially far away

REALISATION OF THE APPROACH Evaluating Competence HITS Web pages are seen as hubs (pages that link to relevant documents, i.e., authorities) and authorities (pages whose content satisfy a query) Intuition: good hubs points to good authorities, and good authorities are pointed to by good hubs J. M. Kleinberg. “Authoritative Sources in a Hyperlinked Environment”. Journal of the ACM, 1999 Swap WWW graph with network of judgements Competent users point to relevant resources, and relevant resources are pointed to by competent users 0.25 0.25 0.25 0.25 0.75 0.25 0.50 0.25 0.75 1.00 1.25 0.75 0.20 0.27 0.33 0.20

Evaluating Competence

HITS

Web pages are seen as hubs (pages that link to relevant documents, i.e., authorities) and authorities (pages whose content satisfy a query)

Intuition: good hubs points to good authorities, and good authorities are pointed to by good hubs

Swap WWW graph with network of judgements

Competent users point to relevant resources, and relevant resources are pointed to by competent users

REALISATION OF THE APPROACH Evaluating Competence HITS suffers from the Tightly Knit Community syndrome SALSA To solve TKC, divide each weight in forward step by node outdegree, and in backward step by authority indegree Side effect: Niche judgements are rewarded more HITS misses on subjectivity Set the only non-zero weight hub being the evaluating node itself At each iteration, stop with probability (1-  ) Low  implies faster convergence and higher subjectivity R. Lempel and S. Moran. “SALSA: the Stochastic Approach to for Link Structure Analysis”. ACM TOIS, 2001

Evaluating Competence

HITS suffers from the Tightly Knit Community syndrome

SALSA

To solve TKC, divide each weight in forward step by node outdegree, and in backward step by authority indegree

Side effect: Niche judgements are rewarded more

HITS misses on subjectivity

Set the only non-zero weight hub being the evaluating node itself

At each iteration, stop with probability (1-  )

Low  implies faster convergence and higher subjectivity

REALISATION OF THE APPROACH Social Filtering Culprit of HITS: backward step (having expressed a judgement is no guarantee of honesty) SOFIA Step 1: Pre-compute PPR on the Web of Trust to compute hubs (nodes) reputation as seen by A Step 2: Start subjective version of HITS/SALSA Step 3: At each backward step, redistribute weights from judgements to users proportionally to users’ reputation as computed by PPR

Social Filtering

Culprit of HITS: backward step (having expressed a judgement is no guarantee of honesty)

SOFIA

Step 1: Pre-compute PPR on the Web of Trust to compute hubs (nodes) reputation as seen by A

Step 2: Start subjective version of HITS/SALSA

Step 3: At each backward step, redistribute weights from judgements to users proportionally to users’ reputation as computed by PPR

REALISATION OF THE APPROACH Parameters: G=(V,E) , V=U U J , reputation vector r computed with PPR on U Returns: vector t’ with ranking of judgements n=size of U, m=size of J t=0 n , t A =1 while (not converged) do {forward step: from users to judgements} t’=0 m for all (u,j) in E do t’ j =t’ j + [w uj / (  k in J w uk )]*t u {backward step: from judgements to users} t=0 n ; t A =1-  ; for all (u,j) in E do t u =t u +  * [(w uj *r u )/(  v in U w vj *r v )]*t’ j end while return t’

Parameters: G=(V,E) , V=U U J , reputation vector r computed with PPR on U

Returns: vector t’ with ranking of judgements

n=size of U, m=size of J

t=0 n , t A =1

while (not converged) do

{forward step: from users to judgements}

t’=0 m

for all (u,j) in E do

t’ j =t’ j + [w uj / (  k in J w uk )]*t u

{backward step: from judgements to users}

t=0 n ; t A =1-  ;

for all (u,j) in E do

t u =t u +  * [(w uj *r u )/(  v in U w vj *r v )]*t’ j

end while

return t’

BROWSING THE TAPESTRY – OUTLINE Philosophy of the Approach Intent Competence Social Filtering Realisation of the Approach Intent: Personalised PageRank Competence: HITS Social Filtering: SOFIA Experiments Accuracy Robustness

Philosophy of the Approach

Intent

Competence

Social Filtering

Realisation of the Approach

Intent: Personalised PageRank

Competence: HITS

Social Filtering: SOFIA

Experiments

Accuracy

Robustness

EXPERIMENTS Datasets Citeseer ( http://citeseer.ist.psu.edu ) Social network: co-autorship data (A and B are connected if they wrote papers together) Judgements: citations (if X cites Y, then the authors of X make the implicit judgement “Y is relevant”) Highly clustered subset of the whole graph 10,000 authors 182,675 papers

Datasets

Citeseer ( http://citeseer.ist.psu.edu )

Social network: co-autorship data (A and B are connected if they wrote papers together)

Judgements: citations (if X cites Y, then the authors of X make the implicit judgement “Y is relevant”)

Highly clustered subset of the whole graph

10,000 authors

182,675 papers

EXPERIMENTS Datasets Last.fm ( http://www.last.fm ) Social network: explicit friends list Judgements: top 50 listened artists chart for each user (implicit judgement “I like to listen to songs by X”) BFS crawl of 10,000 users 51,654 artists

Datasets

Last.fm ( http://www.last.fm )

Social network: explicit friends list

Judgements: top 50 listened artists chart for each user (implicit judgement “I like to listen to songs by X”)

BFS crawl of

10,000 users

51,654 artists

EXPERIMENTS – ACCURACY How to evaluate accuracy? Goal: rank highly (i.e., recommend) judgements that a user would approve Process: Hide a random judgement Run SOFIA If the algorithm performs well, the hidden judgement will have a high ranking Citeseer: guess a missing citation from a paper Last.fm: find a missing artist in a chart

How to evaluate accuracy?

Goal: rank highly (i.e., recommend) judgements that a user would approve

Process:

Hide a random judgement

Run SOFIA

If the algorithm performs well, the hidden judgement will have a high ranking

Citeseer: guess a missing citation from a paper

Last.fm: find a missing artist in a chart

EXPERIMENTS – ACCURACY Citeseer Parameters:  =0.5,  =0.3 for SOFIA;  =0.3 for PPR;  =0.05 for N-SOFIA 1709 115 30 8 2 1 PPR 1136 63 12 3 1 1 N-SOFIA 855 31 4 1 1 1 SOFIA 90 75 50 25 10 5

Citeseer

EXPERIMENTS – ACCURACY Last.fm Parameters:  =0.9,  =0.05 for SOFIA;  =0.5 for PPR;  =0.01 for N-SOFIA 16025 2188 344 66 12 5 PPR 6954 822 157 32 6 2 N-SOFIA 7429 992 174 32 6 2 SOFIA 90 75 50 25 10 5

Last.fm

EXPERIMENTS – ROBUSTNESS How to evaluate robustness to Sybil attacks? Goal: defend against malicious judgements Process: Create coalition of 100 Sybils Pick random victim A All Sybils copy A’s judgements, then add a link to X Study ranking of X before and after the attack , on the victim and on other nodes

How to evaluate robustness to Sybil attacks?

Goal: defend against malicious judgements

Process:

Create coalition of 100 Sybils

Pick random victim A

All Sybils copy A’s judgements, then add a link to X

Study ranking of X before and after the attack , on the victim and on other nodes

EXPERIMENTS – ROBUSTNESS 2583 5 5165 10 38741 25827 12914 No attack - Any 75 50 25 Role k Algorithm

EXPERIMENTS – ROBUSTNESS 2297 1285 334 2583 5 4459 2353 559 5165 10 33322 13371 3101 20493 8757 2012 10730 4759 1092 1 10 100 PPR 38741 25827 12914 No attack - Any 75 50 25 Role k Algorithm

EXPERIMENTS – ROBUSTNESS 1 34 2297 1285 334 2583 5 1 85 4459 2353 559 5165 10 1 3132 1 1185 1 348 Victim Other N-SOFIA 33322 13371 3101 20493 8757 2012 10730 4759 1092 1 10 100 PPR 38741 25827 12914 No attack - Any 75 50 25 Role k Algorithm

EXPERIMENTS – ROBUSTNESS Lower values of  improve attack resilience, at the expense (small) of accuracy 679 2264 41 1082 1 215 1 34 2297 1285 334 2583 5 1386 4409 132 2126 2 391 1 85 4459 2353 559 5165 10 31765 33064 2815 14718 197 5571 11182 19186 1311 8779 74 2649 3406 9599 469 4612 13 1040 Victim Other Victim Other Victim Other 1 10 100 SOFIA 1 3132 1 1185 1 348 Victim Other N-SOFIA 33322 13371 3101 20493 8757 2012 10730 4759 1092 1 10 100 PPR 38741 25827 12914 No attack - Any 75 50 25 Role k Algorithm

UNLOCKING THE TAPESTRY

UNLOCKING THE TAPESTRY Present Paradox Centralised sharing (i.e., Web 2.0) of distributedly-produced & location-meaningful content Risks Scalability Interoperability Usability

Present Paradox

Centralised sharing (i.e., Web 2.0) of distributedly-produced & location-meaningful content

Risks

Scalability

Interoperability

Usability

UNLOCKING THE TAPESTRY Evaluate Browse Build

UNLOCKING THE TAPESTRY HOW TO BUILD AN UNLOCKED TAPESTRY

HOW TO BUILD AN UNLOCKED TAPESTRY Content sharing algorithms and frameworks Impact of mobility With Lucia Del Prete, Liam McNamara & Cecilia Mascolo Impact of uncooperative and malicious users With Afra J. Mashhadi & Sonia Ben Mokhtar

Content sharing algorithms and frameworks

Impact of mobility

With Lucia Del Prete, Liam McNamara & Cecilia Mascolo

Impact of uncooperative and malicious users

With Afra J. Mashhadi & Sonia Ben Mokhtar

UNLOCKING THE TAPESTRY HOW TO BROWSE AN UNLOCKED TAPESTRY

HOW TO BROWSE AN UNLOCKED TAPESTRY Distributed SOFIA-like algorithms Partial knowledge With Matteo Dell’Amico “ Making sense” of content Folksonomy & Ontology With Valentina Zanardi & Sonia Ben Mokhtar

Distributed SOFIA-like algorithms

Partial knowledge

With Matteo Dell’Amico

“ Making sense” of content

Folksonomy & Ontology

With Valentina Zanardi & Sonia Ben Mokhtar

UNLOCKING THE TAPESTRY HOW TO EVALUATE OUR APPROACHES

HOW TO EVALUATE Data Processing How to overlay different datasets about people movement/colocation, interests and social relations in a meaningful way With Afra J. Mashhadi & Sonia Ben Mokhtar Data Gathering Market-based scenario With Daniele Quercia & Liam McNamara

Data Processing

How to overlay different datasets about people movement/colocation, interests and social relations in a meaningful way

With Afra J. Mashhadi & Sonia Ben Mokhtar

Data Gathering

Market-based scenario

With Daniele Quercia & Liam McNamara

THANK YOU! Research Group Website http://mobisys.cs.ucl.ac.uk Research Group Blog http://mobblog.cs.ucl.ac.uk/

Research Group Website

http://mobisys.cs.ucl.ac.uk

Research Group Blog

http://mobblog.cs.ucl.ac.uk/