View
218
Download
0
Embed Size (px)
Citation preview
A Reputation Scheme Combining Trust and Incentive for P2P File
Sharing
A Reputation Scheme Combining Trust and Incentive for P2P File
Sharing
Mao Yang , Qinyuan Feng , Yafei Dai , Zheng Zhang
Peking University, Beijing, ChinaMicrosoft Research Asia, Beijing, China
ICDCS2007 , Toronto , Canada
Mao Yang , Qinyuan Feng , Yafei Dai , Zheng Zhang
Peking University, Beijing, ChinaMicrosoft Research Asia, Beijing, China
ICDCS2007 , Toronto , CanadaLi Xiaoming, Peking University
OutlineOutline
MotivationsMotivations The problemThe problem Proposal for a solutionProposal for a solution An experiment to understand An experiment to understand
some features of the solutionsome features of the solution ConclusionsConclusions
Motivations from two Motivations from two endsends
PhilosophicalPhilosophical Making “to be trusted” be part of incentive for Making “to be trusted” be part of incentive for
living in a p2p worldliving in a p2p world
Technical (practical)Technical (practical) Enrich initial trust-matrix M with richer Enrich initial trust-matrix M with richer
information (for computational accuracy as well information (for computational accuracy as well as efficiency) as efficiency)
Eigen-trust Muilt-trust
The “basics” of P2P file The “basics” of P2P file sharing system sharing system ““Axiom 1”: a p2p file sharing systeAxiom 1”: a p2p file sharing syste
m is in a healthy state if it has enougm is in a healthy state if it has enough number of users and the number dh number of users and the number does not decrease.oes not decrease. in a macro viewin a macro view
““Axiom 2”: the only reason for an iAxiom 2”: the only reason for an individual being a p2p user is to get ndividual being a p2p user is to get what he wants easily.what he wants easily. People tend to be free riders for public goods.People tend to be free riders for public goods. After all, “free riding” is a fundamental phenomena After all, “free riding” is a fundamental phenomena
in human society (Olson, Mancur 1966, The Logic in human society (Olson, Mancur 1966, The Logic of Collective Action. Cambridge University Press ) of Collective Action. Cambridge University Press )
The “basics” of P2P file The “basics” of P2P file sharing system sharing system
““Law 1”: a volunteer-based p2p Law 1”: a volunteer-based p2p system is in a healthy state if it system is in a healthy state if it has enough good resources that has enough good resources that can be easily accessedcan be easily accessed A: A: availabilityavailability (quantity of quality resources) (quantity of quality resources) B: B: accessibilityaccessibility (high bandwidth, easy to find, not (high bandwidth, easy to find, not
confusing, not frustrating)confusing, not frustrating)
““Law 2”: a volunteer-based p2p Law 2”: a volunteer-based p2p system being in a healthy state is system being in a healthy state is not because we have eliminated not because we have eliminated free-riders and colluders, etc.free-riders and colluders, etc.
ButBut ,…,… Selfish usersSelfish users
free-riders: download a lot, upload a little (purposefree-riders: download a lot, upload a little (purposely or not), negative to Aly or not), negative to A
Whitewashers, colluders: try to take advantage of iWhitewashers, colluders: try to take advantage of incentive policy in a p2p system without actual contncentive policy in a p2p system without actual contribution, negative to both A and B.ribution, negative to both A and B.
““Malicious” usersMalicious” users Intentionally upload fake files, negative to BIntentionally upload fake files, negative to B
• To take advantage of traffic-based incentive policy
• To pollute the resource environment, so that innocent users get frustrated.
How to keep the users: the core concern of a p2p system owner
Two general measures to Two general measures to keep userskeep users Incentives (awards) Incentives (awards) improve A improve A
provide some kind of reward for contributions provide some kind of reward for contributions such as points in Maze, TFT in BT and eMulesuch as points in Maze, TFT in BT and eMule
Trust Trust improve B improve B Knowing the trustworthy of users or files, user has Knowing the trustworthy of users or files, user has
more confidence about what he is gettingmore confidence about what he is getting such as EigenTrust such as EigenTrust
A duality formulation of A duality formulation of award and trustaward and trust
AwardAward :: attribute of a attribute of a downloaderdownloader so that a so that a uploaderuploader can use it to diffe can use it to differentiate service to multiple rentiate service to multiple downlodownloadersaders
TrustTrust :: attribute of a attribute of a uploaderuploader so t so that a hat a downloaderdownloader can use it to diffe can use it to differentiate service from multiple rentiate service from multiple uplouploadersaders
Why one needs a high Why one needs a high trust ?trust ?
Award is useful – get better service, so Award is useful – get better service, so one benefit from it directlyone benefit from it directly
Trust is also useful – Trust is also useful – when combined wwhen combined with awardith award – “higher trust – “higher trust higher aw higher award”, so one benefit from it indirectlyard”, so one benefit from it indirectly
In Chinese old saying, “good return foIn Chinese old saying, “good return for good behavior, and bad return for bar good behavior, and bad return for bad behavior”d behavior” (善有善报,恶有恶报) (善有善报,恶有恶报)
Reputation = f Reputation = f (award, trust)(award, trust)
Reputation = f (traffic–based award, trust)Reputation = f (traffic–based award, trust) In Chinese, respect = f (hats, words)In Chinese, respect = f (hats, words)
一个人得到的社会尊重 一个人得到的社会尊重 = f (= f ( 光环,口碑光环,口碑 ))
Global or pair wise ?
Reputation matrix, r(i,j) = fReputation matrix, r(i,j) = fijij(aw(award, trust)ard, trust) Award – traffic basedAward – traffic based :: more upload, hmore upload, h
igher award – easy to measureigher award – easy to measure What about trust (what is good behavioWhat about trust (what is good behavio
r) ?r) ? Vote on files honestly Vote on files honestly Retain good files for long timeRetain good files for long time Remove bad files quicklyRemove bad files quickly Upload good filesUpload good files Rank users honestlyRank users honestly
Granularity: users vs. files; Granularity: users vs. files; Measurement: explicit vs. implicitMeasurement: explicit vs. implicit
File based
Download based
User based
From behavior to trust value From behavior to trust value (( 11 )) Vote on files honestlyVote on files honestly
EEEEikik: an explicit value for file k assigned from user i: an explicit value for file k assigned from user i
Retain good files for long time, and reRetain good files for long time, and remove bad files quicklymove bad files quickly IEIEikik: an implicit value for file k due to action of user i: an implicit value for file k due to action of user i
Combing Combing EEik and IEik to form Eik, the opinion of user i to the file k.
File based trust File based trust relationshiprelationship
Fij: the intersection of files evaluated by Ui and Uj
FTFTijij: file based trust relationship bet: file based trust relationship between users Uween users Uii and U and Ujj
The more they agree on files, the stThe more they agree on files, the stronger trust they have (symmetric). ronger trust they have (symmetric).
If he is a good guy, he probably will behave well. Thus if I want him thinks me good (so I can get things from him easily), I should also behave well.
From behavior to trustFrom behavior to trust (( 22 )) Upload good filesUpload good files
EEikik*S*Skk, the trustworthy of file k considered by user , the trustworthy of file k considered by user
i. Summing over k from one user j, we get the opi. Summing over k from one user j, we get the opinion of user i about user j based on downloadsinion of user i about user j based on downloads
Thus, we will not have simplistic trThus, we will not have simplistic trafic-based award for upload. Ratheafic-based award for upload. Rather, the contribution in terms of volur, the contribution in terms of volume is attributed with trust.me is attributed with trust.
Download volume combined Download volume combined with trustwith trust
For all files downloaded from j
Normalized with all users
From behavior to trust From behavior to trust (3)(3)
Rank users Rank users honestlyhonestly
UTij: Ui’s rank of Uj
Uall: all the users
An integrated one-step trust matrix (TM) :
TM(i,j): how much i trusts j based on the three factors
From trust to reputation From trust to reputation (multi trust)(multi trust)
nTMRM We’d like to know the (non We’d like to know the (non
zero) density of TM with zero) density of TM with respect to the probability of file respect to the probability of file evaluationsevaluations Higher the density, lower the nHigher the density, lower the n Also note that density of TM is higher than that of Also note that density of TM is higher than that of
FM FM
The emulation based on The emulation based on Maze logsMaze logs Logs of one month: 103K users, Logs of one month: 103K users,
24M download/upload actions, 24M download/upload actions, 395K files395K files <t, i, j, f>: at time t, i downloaded file f from j <t, i, j, f>: at time t, i downloaded file f from j
Following the time sequence of the Following the time sequence of the logslogs Set a=0, b=0 /** we watch for a/bSet a=0, b=0 /** we watch for a/b For each observed log <t, i, j, f> For each observed log <t, i, j, f>
• b++• if there exists file that has been evaluated by both i and j
before t, then a++, else evaluate (vote) file f with probability k% on behave of i.
For each end of day, record a/b For each end of day, record a/b
Experiments for different k’Experiments for different k’ss
Note, this is only an approximation of FM density
The use of reputation The use of reputation valuesvalues
GGikik: goodness of file k reported to user i, base: goodness of file k reported to user i, based on collective evaluations and reputationsd on collective evaluations and reputations
KKuu: the set of users that have evaluated k: the set of users that have evaluated k
Fake files identification. If i Fake files identification. If i wants to download file k.wants to download file k.
To get “true” value of Gik,i should try to make RMij as true as possible
Naturally, it can also be used as priority for downloading
ConclusionsConclusions
We proposed a pair wise We proposed a pair wise reputation scheme combining reputation scheme combining trust and incentivetrust and incentive philosophically, encouraging user to behave well philosophically, encouraging user to behave well
in a file sharing system for his own goodin a file sharing system for his own good technically, richer direct trust relation for accuracy technically, richer direct trust relation for accuracy
and ease of computation.and ease of computation.
Some of the possible attack and Some of the possible attack and implementation issues are also implementation issues are also discussed in the paper. discussed in the paper.
Future worksFuture works
Fine tune of the parameters with mFine tune of the parameters with more emulationsore emulations
More direct relation between “gooMore direct relation between “good behavior” (especially those subjd behavior” (especially those subjective behavoir) and reputation valective behavoir) and reputation valuesues How to let the “quality of ith row” be reflected on How to let the “quality of ith row” be reflected on
the ith colunm ?the ith colunm ?
Deployment on real systemsDeployment on real systems