differenti dprivat network trace analysi abstract potenti network trace analysi provid guarante differenti privaci differenti privaci provabl obscur presenc absenc individu record dataset major limit analys present express higher level declar languag analysi result random return analyst report experi conduct divers set analys differenti privat manner express target analys approxim express requir error level low run analys real dataset find error introduc sake privaci low high level privaci factor learn toolkit analys conclud differenti privaci show promis broad class network analys categori subject descriptor comput commun network miscellan data dparallel toolkit inform retriev differenti privat combinatori optim problem metric space point client select set facil locat minim averag distanc client nearest facil studi median problem approxim algorithm hard result note object function encourag open facil area client solut good idea client locat rais quandari locat client sensit inform privat design good algorithm problem preserv privaci client paper initi systemat studi algorithm discret optim problem framework differenti privaci formal idea protect privaci individu input element show problem good approxim algorithm preserv differenti privaci case imposs preserv cryptograph definit privaci comput trivial approxim optim solut entir solut median problem problem vertex set cover min cut median facil locat steiner tree give approxim algorithm lower bound problem recent introduc submodular maxim problem combinatori public project cpp shown papadimitri al inapproxim subpolynomi multipl factor effici truth algorithm give differenti privat approxim truth algorithm achiev logarithm addit approxim simpl practic algorithm differenti privat data releas privaci integr queri extens platform privaci dpreserv data analysi differenti privat recommend system build privaci netflix prize contend problem produc recommend collect user behavior simultan provid guarante privaci user specif netflix prize data set lead algorithm adapt framework differenti privaci unlik prior privaci work concern cryptograph secur comput recommend differenti privaci constrain comput preclud infer underli record output algorithm necessarili introduc uncertainti nois comput trade accuraci privaci find lead approach netflix prize competit adapt provid differenti privaci significantli degrad accuraci adapt algorithm explicitli factor part aggreg learn phase perform differenti privaci guarante individu recommend phase learn correl individu data provid person recommend adapt trivial involv care analysi record sensit algorithm calibr nois post process step mitig impact nois measur empir trade accuraci privaci adapt find provid trivial formal privaci guarante outperform cinematch baselin netflix special issu thirti deighth annual acm symposium theori comput stoc comput inform retriev perform measur effici presenc ti score abstract inform retriev commun varieti perform measur evalu effect score function paper show adapt popular measur precis recal averag precis reciproc rank normal discount cumul gain cope score function assign ti score result search ti score impos partial order result mean multipl order result set perform differ approach cope ti averag perform valu result order gener result permut requir super exponenti time approach present paper comput precis perform approach averag permut effici origin tie oblivi measur data collect denforc privaci pollster wish collect privat sensit data number distrust individu pollster convinc respond trustworthi altern mechan respond insist ensur mismanag data detect publicli demonstr detail problem provid simpl data submiss protocol properti leakag privat data pollster result evid transgress evid fabric break cryptograph assumpt guarante respons pollster post privaci bond forfeit provid evid leakag respond assur penalti appli leaki pollster protect spuriou indict ensur honest pollster disincent particip scheme decentr algorithm spectral analysi abstract abil robot team reconfigur applic metamorph robot chang shape swarm motion goal biolog system avoid predat mobil buoi clean oil spill situat auxiliari constraint connect team member limit maximum hop count satisfi reconfigur paper show estim control graph connect accomplish decentr manner describ decentr estim procedur agent track algebra connect time vari graph base estim propos decentr gradient control agent maintain global connect motion mechan design differenti privaci studi role privaci preserv algorithm prevent leakag specif inform particip plai design mechan strateg agent encourag player honestli report inform specif show recent notion differenti privaci addit intrins virtu ensur particip limit effect outcom mechan consequ limit incent lie precis mechan differenti privaci approxim domin strategi arbitrari player util function automat resili coalit easili repeat studi special case unlimit suppli auction problem provid result digit good auction attribut auction auction arbitrari structur constraint price import prelud develop privaci preserv auction mechan introduc studi gener previou privaci work accommod high sensit auction set singl particip dramat alter optim fix price slight chang offer price revenu optim privaci parti px hold bit input privat comput px learn addit inform inform theoret sens input function similarli learn addit inform paper basic question theori privat comput function privat comput commun complex protocol privat comput function case protocol exist complet combinatori character privat comput function character deriv tight bound round complex privat comput function design optim privat protocol comput function show delta function privat fast comput low drank matrix approxim practic applic matrix interest approxim low rank introduc techniqu exploit spectral structur acceler orthogon iter lanczo iter common method comput approxim techniqu amount independ sampl quantiz entri input matrix speed comput reduc number entri length represent analysi base observ sampl quantiz view ad random matrix entri independ random variabl bound varianc random matric poss signic linear structur prove eect sampl quantiz vanish low rank approxim comput fact promin linear structur data aord discard ultim faster discov give bound qualiti approxim frobeniu norm author address microsoft research microsoft redmond wa author address depart comput scienc univers washington seattl wa research perform visit microsoft research support part microsoft corpor data collect denforc privaci pollster wish collect privat sensit data number distrust individu pollster convinc respond trustworthi altern mechan respond insist ensur mismanag data detect publicli demonstr detail problem provid simpl data submiss protocol properti leakag privat data pollster result evid transgress evid fabric break cryptograph assumpt guarante respons pollster post privaci bond forfeit provid evid leakag respond assur penalti appli leaki pollster protect spuriou indict ensur honest pollster disincent particip scheme data privaci distribut nois gener abstract work provid effici distribut protocol gener share random nois secur malici particip purpos nois gener creat distribut implement privaci preserv statist databas recent paper databas privaci obtain perturb true answer databas queri addit small amount gaussian exponenti distribut random nois comput power evenasimpl form databas queryi form di sum row databas function appli data row demonstr distribut implement elimin trust databas administr result nois gener independ interest gener gaussian nois introduc techniqu distribut share unbias coin fewer execut verifi secret share need previou approach reduc afactorofn gener exponenti distribut nois shallow circuit gener arbitrarili ident bias coin amort cost unbias random bit apiec independ bia combin bit bias obtain exponenti distribut calibr nois sensit privat data analysi abstract continu line research initi privacypreserv statist databas trust server hold databas sensit inform queri function map databas real call true answer result appli databas protect privaci true answer perturb addit random nois gener carefulli chosen distribut respons true answer nois return user previou work focus case noisi sum xi xi denot ith row databas map data base row extend studi gener function prove privaci preserv calibr standard deviat nois sensit function roughli speak amount singl argument chang output analysi show applic substanti nois need previous understood case step clean character privaci term indistinguish transcript addition obtain separ result show increas interact sanit mechan interact introduct continu line research initi privaci statist data base statist quantiti comput sampl intuit databas repres sampl underli popul goal ofa privaci preserv statist databas enabl user learn properti popul protect privaci individualcontributor spectral learn mixtur distribut problem learn mixtur distribut spectral method deriv tight character method specif mixtur sampl denot empir covari matrix mix weight compon prove simpl algorithm spectral project singl linkag cluster properli classifi point sampl separ term depend concentr properti distribut mixtur term small distribut includ gaussian log concav result bound learn mixtur arbitrari gaussian term requir separ hand prove mean mix weight set matric separ appli spectral project gaussian mixtur collaps complet mean covari matric project mixtur ident peer dto dpeer worm threat defens peer peer worm exploit common vulner member host network spread topolog network potenti effect strategi random scan locat victim paper describ danger pose worm initi studi mitig mechan paper explor feasibl defens infrastructur insid network outlin challeng evalu defens mechan worm reveal correl contain overlai topolog network experi suggest number design direct improv resili network worm attack practic privaci sulq framework statist databas trust administr introduc nois queri respons goal maintain privaci individu databas entri databas queri consist pair set row databas function map databas row true answer di noisi version releas respons queri result dinur dwork nissim show strong form privaci maintain surprisingli small amount nois sampl error provid total number queri sublinear number databas row call queri slightli noisi repli sulq linear queri primit assumpt sublinear reason databas grow increasingli larg extend work wai modifi privaci analysi real valu function arbitrari row type consequ greatli improv bound nois requir privaci examin comput power sulq primit show power slightli noisi version comput carri invoc primit princip compon analysi mean cluster perceptron algorithm id algorithm appar algorithm oper statist queri learn model privaci public databas abstract initi theoret studi censu problem inform censu individu respond give privat inform trust parti censu bureau publish sanit version data fundament conflict requir privaci respond util sanit data unlik studi secur function evalu privaci preserv extent specif function goal censu problem privaci paramount intuit thing learn safe learn import contribut work definit privaci privaci compromis statist databas method describ compar privaci offer specif sanit techniqu obtain privaci result sanit techniqu show combin cross train obtain util result involv cluster profit dmaxim envi dfree price stackelberg price problem direct uncapacit network tariff defin oper leader subset arc tariff arc cost arc assum client follow rout demand independ path minim total cost problem find tariff maxim oper revenu motiv problem telecommun network restrict version problem assum client util oper tariff arc problem equival price bridg client order cross river prove problem apx hard show uniform price yield approxim ln approxim upper bound total demand client discuss polynomi solvabl special case present short comput studi instanc franc addit problem addit restrict oper serv client prove problem admit approxim algorithm reason perform guarante np zpp prove exist approxim algorithm privaci public databas abstract initi theoret studi censu problem inform censu individu respond give privat inform trust parti censu bureau publish sanit version data fundament conflict requir privaci respond util sanit data unlik studi secur function evalu privaci preserv extent specif function goal censu problem privaci paramount intuit thing learn safe learn import contribut work definit privaci privaci compromis statist databas method describ compar privaci offer specif sanit techniqu obtain privaci result sanit techniqu show combin cross train obtain util result involv cluster privaci dpreserv histogram censu individu respond give privat inform trust parti censu bureau publish sanit version data fundament conflict requir privaci respond util sanit data note framework inher noninteract recent chawla al tcc initi theoret studi censu problem present intuit appeal definit privaci breach call isol formal specif requir data sanit algorithm access sanit data increas adversari abil isol individu show data drawn uniformli highdimension hypercub recurs histogram sanit preserv privaci high probabl extend result wai develop method comput privaci preserv histogram sanit round distribut uniform distribut high dimension ball sphere problem challeng unlik hypercub natur histogram distribut long thin cell hurt proof privaci develop techniqu random histogram construct hypercub hyperspher permit appli result approxim quantiti interest cost minimum span tree cost optim solut facil locat problem data point histogram count privaci preserv fashion uniform approach acceler pagerank comput note simpl reformul tradit power iter algorithm comput stationari distribut markov chain commun current probabl valu neighbor step node commun probabl reformul enabl larg degre flexibl manner node updat valu lead arrai optim featur includ faster converg cient increment updat robust distribut implement spectral analysi random graph skew degre distribut extend spectral method random graph skew degre distribut degre base normal close connect normal laplacian normal base intuit drawn perturb theori random matric effect boost expect random adjac matrix increas varianc entri lead perturb bound primari implic result li realm spectral analysi random graph skew degre distribut ubiquit power law graph recent mihail papadimitri argu randomli gener graph satisfi power law degre distribut spectral analysi adjac matrix simpli produc neighborhood high degre node eigenvector miss embed structur present gener model incorpor latent structur prove appli transform spectral analysi succe recov latent structur high probabl decentr algorithm spectral analysi abstract abil robot team reconfigur applic metamorph robot chang shape swarm motion goal biolog system avoid predat mobil buoi clean oil spill situat auxiliari constraint connect team member limit maximum hop count satisfi reconfigur paper show estim control graph connect accomplish decentr manner describ decentr estim procedur agent track algebra connect time vari graph base estim propos decentr gradient control agent maintain global connect motion web search hub synthesi present model web search captur unifi manner critic compon problem link structur web gener content web document gener human searcher gener queri kei unif li captur correl compon term proxim share latent semant space combin model correct answer search queri defin evalu web search algorithm rigor present web search algorithm base spectral techniqu prove guarante produc approxim correct answer model algorithm assum knowledg model defin model accuraci spectral partit random graph problem bisect graph color cliqu gener believ hard worst case solv input data drawn randomli distribut graph accept solut paper show simpl spectral algorithm solv problem averag case gener problem partit graph base edg densiti case approach meet exce previou paramet introduc substanti gener appli spectral techniqu foremost observ problem expect adjac matrix low rank matrix structur solut evid sampl techniqu kernel method propos random techniqu speed kernel princip compon analysi level sampl quantiz gram matrix train random round evalu kernel expans random project evalu kernel case give sharp bound accuraci obtain approxim fast comput low rank matrix practic applic matrix interest approxim low rank introduc techniqu exploit spectral structur acceler orthogon iter lanczo iter common method comput approxim techniqu amount independ sampl quantiz entri input matrix speed comput reduc number entri length represent analysi base observ sampl quantiz view ad random matrix entri independ random variabl bound varianc random matric poss signic linear structur prove eect sampl quantiz vanish low rank approxim comput fact promin linear structur data aord discard ultim faster discov give bound qualiti approxim frobeniu norm author address microsoft research microsoft redmond wa author address depart comput scienc univers washington seattl wa research perform visit microsoft research support part microsoft corpor spectral analysi data 