application-specific secure gathering of consumer preferences … · 2016-09-26 · mobile access...
TRANSCRIPT
Application-SpecificSecureGatheringofConsumerPreferencesandFeedbackin
Information-CentricNetworks
ComputerScienceDepartmentNewMexicoStateUniversity
RezaTourani,Satyajayant (Jay)Misra,TravisMick
NewMexicoStateUniversity,NM
Outline
v IntroductionandMotivationv ProblemDefinitionv RequirementsandPreliminariesv FeedbackCollectionandDeliveryApproachesv ConclusionsandFutureWork
NewMexicoStateUniversity,NM
Outline
v IntroductionandMotivationv ProblemDefinitionv RequirementsandPreliminariesv FeedbackCollectionandDeliveryApproachesv ConclusionsandFutureWork
NewMexicoStateUniversity,NM
NewMexicoStateUniversity,NM
Clientminingiswidespread.
NewMexicoStateUniversity,NM
Benefitsofclientminingandrecommendersystems.
Influenceon80%ofhoursstreamedatNetflix(2016)
Approximately 35%increaseinAmazonrevenue(2013)
50%ofLinkedInjobapplicationsandjobviewsbymembers(2011)
NewMexicoStateUniversity,NM
AmazonCloud
CDNs
NetflixDataCenter(www.netflix.com)
ClientNewuserregistrationUseraccountbilling
NetflixCommunicationFlow
Redirectingusers
NetflixServer
AuthenticationManifestFilePeriodicUpdates
CDNRoutingDRM
NewMexicoStateUniversity,NM
www.hulu.com(HostedbyAkamai)
Client
Client-ServerInteractioninHulu
s.hulu.com(HostedbyAkamai)
t.hulu.com(HostedbyHuku)
3CDNs
Outline
v IntroductionandMotivationv ProblemDefinitionv RequirementsandPreliminariesv FeedbackCollectionandDeliveryApproachesv ConclusionsandFutureWork
NewMexicoStateUniversity,NM
RequestflowinICNmulti-levelarchitecture.
NewMexicoStateUniversity,NM
NewMexicoStateUniversity,NM
DataflowinICNmulti-levelarchitecture.
Pervasivecachingeliminatescontactingproviderforpopularcontent.
NewMexicoStateUniversity,NM
CacheHit!
Howtotrackclientwithoutcommunication?
NewMexicoStateUniversity,NM
Cachingunderminesgatheringofaccessstatistics.
Outline
v IntroductionandMotivationv ProblemDefinitionv RequirementsandPreliminariesv FeedbackCollectionandDeliveryApproachesv ConclusionsandFutureWork
NewMexicoStateUniversity,NM
NewMexicoStateUniversity,NM
ICNrequirementsforsuccessfulclientmining.
Securefeedbackcollection
ContentProvider
Independent
PreservingUserPrivacy
PreciseStatistics
NewMexicoStateUniversity,NM
ContentCategorization
GeneratedBeforehand
PubliclyAvailable
RequireAccessControl
GeneratedOn-Demand
StaticContent
DynamicContent
PrivateContent
PublicContent
• Generated in advance• Require access control• Cacheable
• Generated in advance• Available publicly• Cacheable
• Generated by request• Available publicly
• Generated by request• Require access control• Non-Cacheable
NewMexicoStateUniversity,NM
Static-Publicisthelargestcontentcategory.
66%
34%
ContentTypeinNorthAmerica
Static-Public Other
64%
36%
MobileAccessTrafficinNorthAmerica2016
Encrypted Un-Encrypted
NewMexicoStateUniversity,NM
Abiggerportionofmobileaccesstrafficisencryptedincomparisontofixedaccesstraffic.
29%
71%
FixedAccessTrafficinNorthAmerica2015
Encrypted Un-Encrypted
37%
63%
FixedAccessTrafficinNorthAmerica2016
Encrypted Un-Encrypted
Spotlight: Encrypted Internet traffic (https://www.sandvine.com/trends/global-internet-phenomena)
Outline
v IntroductionandMotivationv ProblemDefinitionv RequirementsandPreliminariesv FeedbackCollectionandDeliveryApproachesv ConclusionsandFutureWork
NewMexicoStateUniversity,NM
FeedbackCollectionandDelivery
NewMexicoStateUniversity,NM
PreferenceTracking Mechanisms
Manifest-Free
CollectionbyIntermediate
routers
CollectionbyClients
CollectionbyISP'sServer
Manifest-Based
ManifestfromProvider
ManifestfromISP'sServer
DrawbacksofCollectionbyIntermediateRouters
NewMexicoStateUniversity,NM
CollectionEvent
Per-Interest Per-Hit
Drawbacks
RedundantStatistics
Coarse-levelStatistics
ComputationOverhead
LackofClientID
Drawbacks
Coarse-levelStatistics
ComputationOverhead
LackofClientID
DrawbacksofCollectionbyClients
NewMexicoStateUniversity,NM
Approaches
Content Partitioning AccessControlEnforcement
Drawbacks
UnknownPartition Size
PartitionPublication
Dependency onOnlineServer
CommunicationOverhead
Drawbacks
SuitableforPrivateContent
Dependency onOnlineServer
CommunicationOverhead
CollectionbytheISP’sDesignatedServer
NewMexicoStateUniversity,NM
Benefits
ReducedLatency
CacheUtilization
Independentof Provider
Drawbacks
ISP-ProviderInteraction
InaccurateStatistics
Provider
• OffloadDecryptionKeyorContentPartition
ISP'sServer
• StoresStatistics• ReturnsRequestedKeyorContent
User
• RequestDecryptionKeyorContentPartition
Manifest-BasedApproaches
NewMexicoStateUniversity,NM
ManifestDelivery
Provider(Un-cacheable)
ISP’sServer(Cacheable)
Drawbacks
ExtraLatency
Provider’sAvailability
Un-cachedContent
Drawbacks
SinglePointofFailure
NetworkBottleneck
Outline
v IntroductionandMotivationv ProblemDefinitionv RequirementsandPreliminariesv FeedbackCollectionandDeliveryApproachesv ConclusionsandFutureWork
NewMexicoStateUniversity,NM
NewMexicoStateUniversity,NM
Howaboutevaluation?
vManifest-based approaches scale much better than the otherschemes.
v Communication overhead as means of evaluating efficacy ofthe approaches – Manifest based approaches scale better.
vManifest-based approaches introduce fixed amount ofoverhead per content and the amortized cost will be low.
v There is a theoretical upper bound on the requiredcommunication overhead per content:
Overhead = 𝐷𝑖𝑠𝑡𝑎𝑛𝑐𝑒𝑐𝑙𝑖𝑒𝑛𝑡 − 𝑠𝑒𝑟𝑣𝑒𝑟✕𝑀𝑎𝑛𝑖𝑓𝑒𝑠𝑡𝑠𝑖𝑧𝑒
v Direct interaction between client and providerwith/without help of routers inaccurate and non-scalable.
v A viable feedback collection mechanism should leveragecaching.
vManifest-based feedback collection approaches aremore scalable, especially if it involves infrastructure atthe ISP.
v Comprehensive evaluation of manifest basedapproaches (Provider vs. ISP server) and identify whichapproach in the other class comes closer.
NewMexicoStateUniversity,NM
ConclusionsandFutureWork
Thankyou!Email:[email protected]
NewMexicoStateUniversity,NM
ResearchfundedbytheUSNationalScienceFoundationandtheUSDept.ofDefense.