Tight on budget? tight bounds for r-fold approximate differential privacy
52 pages - 1.3 MB
By s meiser · 2017 · cited by 31 — our method is useful for deriving tight bounds for classical differential privacy mechanisms but can also be applied to any privacy analysis ...
Document in text mode:
TightonBudget?TightBoundsforr-FoldApproximateDifferentialPrivacySebastianMeiser1,EsfandiarMohammadi21UniversityCollegeLondon,UnitedKingdom,e-mail:[email protected],Switzerland,e-mail:[email protected]sequallycontributedtothiswork.September5,2018AbstractManyapplications,suchasanonymouscommunicationsystems,privacy-enhancingdatabasequeries,orprivacy-enhancingmachine-learningmethods,requirerobustguaranteesunderthousandsandsome-timesmillionsofobservations.Thenotionofr-foldapproximatedifferentialprivacy(ADP)offersawell-establishedframeworkwithaprecisecharacterizationofthedegreeofprivacyafterrobservationsofanattacker.However,existingboundsforr-foldADParelooseand,ifusedforestimatingtherequireddegreeofnoiseforanapplication,canleadtoover-cautiouschoicesforperturbationrandomnessandthustosuboptimalutilityoroverlyhighcosts.Wepresentanumericalandwidelyapplicablemethodforcapturingtheprivacylossofdifferentiallyprivatemechanismsundercomposition,whic...