Order Preserving Encryption for Numeric Data Rakesh Agrawal Jerry Kiernan Ramakrishnan Srikant Yirong Xu IBM Almaden Research Center.

Order Preserving Encryption Order Preserving Encryption for Numeric Datafor Numeric Data

Rakesh AgrawalRakesh AgrawalJerry KiernanJerry Kiernan

Ramakrishnan SrikantRamakrishnan SrikantYirong XuYirong Xu

IBM Almaden Research CenterIBM Almaden Research Center

OutlineOutline

Motivation and IntroductionMotivation and Introduction OPES encryptionOPES encryption Modeling the distributionModeling the distribution Experimental evaluationExperimental evaluation

MotivationMotivation

Encryption is rapidly becoming a Encryption is rapidly becoming a requirement in a myriad of business requirement in a myriad of business settings (e.g., health care, financial, settings (e.g., health care, financial, retail, government), driven by retail, government), driven by legislations (e.g. SB1386, HIPAA)legislations (e.g. SB1386, HIPAA)

Encrypting databases unleashes a host Encrypting databases unleashes a host of problems:of problems:– Performance slowdownPerformance slowdown– Incompatibility with standard database Incompatibility with standard database

featuresfeatures E.g. comparison predicates and the use of indexesE.g. comparison predicates and the use of indexes

– Changes to applications for encryptionChanges to applications for encryption Encryption functions now appear in queriesEncryption functions now appear in queries

if (p1 < p2) then (c1 < c2)if (p1 < p2) then (c1 < c2)

Order Preserving Order Preserving Encryption FunctionEncryption Function

E is an order preserving encryption function,and p1 and p2 are two plaintext values, and

c1 = E(p1)c2 = E(p2)

Threat ModelThreat Model

The storage system used by the DBMS is The storage system used by the DBMS is untrusted, i.e. vulnerable to compromiseuntrusted, i.e. vulnerable to compromise

The DBMS software is trustedThe DBMS software is trusted Ciphertext only attackCiphertext only attack

– The adversary has access to all (but only) The adversary has access to all (but only) encrypted values encrypted values

Guard against percentile exposureGuard against percentile exposure– An adversary should not be able to get even An adversary should not be able to get even

an estimate of true valuesan estimate of true values

Design GoalsDesign Goals

Query results from OPES will be Query results from OPES will be sound and completesound and complete

Comparison operations will be Comparison operations will be performed without decrypting the performed without decrypting the operandsoperands

Standard database indexes can Standard database indexes can be used over encrypted databe used over encrypted data

Tolerate updatesTolerate updates

Integration of Integration of Encryption and Query Encryption and Query

ProcessingProcessing

Select name from Emp where sal > 100000

Queries

Select decrypt (“xsxx”)from “cwlxss”where “xescs” > OPESencrypt(100000)

DBMS

Translation layer

Encrypted dataAnd metadata

Users have a plaintext view of an encrypted database

Plaintext queries are translated into equivalent queries over encrypted data

Tables are encrypted using standard as well as order

preserving encryption

Comparison operators are directly applied over encrypted columns

We hereafter strictly focus on the OPES

algorithms

OutlineOutline


ApproachApproach

Plaintext data has unknown Plaintext data has unknown distributiondistribution

User selects the target User selects the target (ciphertext) distribution(ciphertext) distribution

Ciphertext values exhibit the Ciphertext values exhibit the target distributiontarget distribution

Effect of OPES Effect of OPES Encryption on Encryption on

Plaintext DistributionsPlaintext Distributions

Input: Gaussian, Target: Zipf

Original EncryptedTarget

0

500

1000

1500

2000

2500

3000

Scaled Domain

No. of Poin

ts

Input: Uniform, Target: Zipf

0

500

1000

1500

2000

2500

3000

Scaled Domain

No. of Poin

ts

OPES Key GenerationOPES Key Generation

Sample of source values from the plaintext distribution

Sample of source values from the plaintext distribution

Sample of target values from the ciphertext distribution

Sample of target values from the ciphertext distribution

OPES Key Generation

OPES Key

OPES KeysOPES Keys

SourceUniform Uniform

TargetSource to uniform

Target to uniform

Two Step EncryptionTwo Step Encryption

Source (plaintext) to uniformSource (plaintext) to uniform Uniform to target (ciphertext)Uniform to target (ciphertext)

OPES EncryptionOPES Encryption

SourceUniform Uniform

Target

Encrypt

Decrypt

Step I Step II

Step II

Step I

OutlineOutline


Modeling the Modeling the DistributionDistribution HistogramsHistograms

– Equi-depth, equi-width, waveletsEqui-depth, equi-width, wavelets Number of buckets required unreasonably largeNumber of buckets required unreasonably large Over fitting the modelOver fitting the model

ParametricParametric– Poor estimation for irregular distributionsPoor estimation for irregular distributions

Hybrid [Konig and Weikum 99]Hybrid [Konig and Weikum 99]– Query result size estimationQuery result size estimation– ApproachApproach

Partition the data into bucketsPartition the data into buckets Model the distribution within a bucket as a splineModel the distribution within a bucket as a spline Fixed number of bucketsFixed number of buckets

Our ApproachOur Approach

Hybrid [Konig and Weikum 99]Hybrid [Konig and Weikum 99]– Partition the data into bucketsPartition the data into buckets– Model the distribution within each Model the distribution within each

bucket as a linear splinebucket as a linear spline The number of buckets is not The number of buckets is not

fixedfixed We use MDL to determine the We use MDL to determine the

number of bucket boundariesnumber of bucket boundaries

MDLMDL

The best model for encoding data The best model for encoding data minimizes the sum of the cost ofminimizes the sum of the cost of– Describing the modelDescribing the model– Describing data in terms of the Describing data in terms of the

modelmodel

Model CostsModel Costs

Data CostData Cost– Using a mapping M from [pUsing a mapping M from [pll,p,phh) to [f) to [fll,f,fhh), ),

the cost of encoding pthe cost of encoding pii is is C(pC(pii)=log(f)=log(fii-E(i))-E(i)) DC(pDC(pll,p,phh) = C(p) = C(pll)+C(p)+C(pl+1l+1)+…+C(p)+…+C(ph-1h-1))

Incremental Model CostIncremental Model Cost– Fixed cost for each additional bucketFixed cost for each additional bucket

Boundary valueBoundary value Boundary parametersBoundary parameters

– SlopeSlope– Scale factorScale factor

Computing BoundariesComputing Boundaries

Growth phaseGrowth phase– [p[pll,p,phh) with h-l-1 sorted points {p) with h-l-1 sorted points {pl+1l+1,p,pl+2l+2,…,p,…,ph-1h-1}}

Compute spline for [pCompute spline for [pll,p,phh) ) Compute [fCompute [fll,f,fhh) using the spline) using the spline

– Find further split point pFind further split point pss with f with fss having the having the maximum deviation from the expected valuemaximum deviation from the expected value

Prune phasePrune phase– LB(pLB(pll,p,phh)=DC(p)=DC(pll,p,phh)-DC(p)-DC(pll,p,pss)-DC(p)-DC(pss,p,phh)-IMC)-IMC– GB(pGB(pll,p,phh)=LB(p)=LB(pll,p,phh)+GB(p)+GB(pll,p,pss)+GB(p)+GB(pss,p,phh))– if (GB > 0), the split at pif (GB > 0), the split at pss is retained is retained

ScalingScaling

Source

Uniform

xx xx x

x x x x x

b-1 b b+1

Number of values in a bucket may be disproportional to the size of the bucket

UpdatesUpdates

The scale factor ensures that The scale factor ensures that each distinct plaintext value each distinct plaintext value maps to distinct ciphertext valuesmaps to distinct ciphertext values

Encrypted values need not be Encrypted values need not be recomputed unless the recomputed unless the distribution of plaintext values distribution of plaintext values changeschanges

Quality of EncryptionQuality of Encryption

KS Statistical TestKS Statistical Test– Can we disprove, to a certain Can we disprove, to a certain

required level of significance, the required level of significance, the null hypothesis that two data sets null hypothesis that two data sets are drawn from the same are drawn from the same distribution function?distribution function?

– If not, then the ciphertext If not, then the ciphertext distribution cannot be distinguished distribution cannot be distinguished from the specified target distributionfrom the specified target distribution

DuplicatesDuplicates

AssumptionsAssumptions– A large number of duplicates may leak A large number of duplicates may leak

information about the distribution of information about the distribution of valuesvalues

Alternatively,Alternatively,– Map duplicates to distinct valuesMap duplicates to distinct values– if (f = M(p), f’ = M(p+1))if (f = M(p), f’ = M(p+1))

[f,f’) = M(p)[f,f’) = M(p)– Equality expressed as a rangeEquality expressed as a range– Equi-joins can no longer be expressedEqui-joins can no longer be expressed

However, many numeric attributes (e.g., However, many numeric attributes (e.g., salarysalary) ) may rarely be used in joinsmay rarely be used in joins

OutlineOutline


Experimental Experimental EvaluationEvaluation Percentile exposurePercentile exposure UpdatabilityUpdatability Key sizeKey size Time overheadTime overhead

DatasetsDatasets

CensusCensus– UCI KDD archive, PUMS census data UCI KDD archive, PUMS census data

(30,000) records(30,000) records GaussianGaussian ZipfZipf UniformUniform

Source:Source: GaussianGaussian

Target:Target: ZipfZipf

DefaultDefault

Percentile ExposurePercentile Exposure

Source Source distributiodistributionn

Target Target distributiodistributionn

Average Average change in change in percentilepercentile

CensusCensus GaussianGaussian 3737

CensusCensus ZipfZipf 77

CensusCensus UniformUniform 3838

GaussianGaussian ZipfZipf 4545

GaussianGaussian UniformUniform 1717

ZipfZipf UniformUniform 4444

Time to the Build Time to the Build ModelModel

0

0.1

0.2

0.3

0.4

0.5

10K 100K 1M 10M

Dataset size

Tim

e p

er

tuple

(m

sec.

)

ZipfUniformGaussian

Insertion OverheadInsertion Overhead

0

20

40

60

80

100

1 10 100

1000

10,000

Tuples inserted

% o

verh

ead

10K100K1M10M

Cost of Additional Cost of Additional InsertionInsertion

0

2

4

6

8

10

1 10 100

1000

10,000

Additional tuples inserted in 10M tuple relation

Tim

e p

er

tuple

(m

sec.

)

PlaintextCiphertext

Retrieval OverheadRetrieval Overhead

0

20

40

60

80

100

1 100 10K 1M

Tuples retrieved

% o

verh

ead

10K100K1M10M

Retrieval TimeRetrieval Time

0.001

0.01

0.1

1

10

1

10

100

1000

10K

100K

1M

Tuples retrieved (10M tuple relation)

Tim

e p

er

tuple

(m

sec.

)

PlaintextCiphertext

Related WorkRelated Work

Polynomial functionsPolynomial functions– Ignores the distribution of plaintext/ciphertext valuesIgnores the distribution of plaintext/ciphertext values

Database as a serviceDatabase as a service– Requires post processing of query resultsRequires post processing of query results

Privacy homomorphismsPrivacy homomorphisms– Comparison operations not investigatedComparison operations not investigated

Keyword searches on encrypted dataKeyword searches on encrypted data– Designed for keyword retrievalDesigned for keyword retrieval– Range queries not supportedRange queries not supported

Smartcard-based schemesSmartcard-based schemes– Infeasible for large rangesInfeasible for large ranges

Order-preserving hashingOrder-preserving hashing– Protecting the hash values from cryptanalysis is not a Protecting the hash values from cryptanalysis is not a

concern, nor is deciphering plaintext values from hash concern, nor is deciphering plaintext values from hash valuesvalues

– Designed for static collectionsDesigned for static collections

Closing RemarksClosing Remarks

Ensuring safety without impeding the Ensuring safety without impeding the flow of information is a hard problemflow of information is a hard problem

Current choicesCurrent choices– Plaintext databasePlaintext database– Encrypted databases with loss of Encrypted databases with loss of

functionality or performancefunctionality or performance Our approach focused on the trade-off Our approach focused on the trade-off

between security and efficiencybetween security and efficiency We developed an algorithm which We developed an algorithm which

could easily be integrated with current could easily be integrated with current systemssystems

BackupBackup

EncodeEncode

Encode(p) = z(spEncode(p) = z(sp22+p)+p)

p p [0,p [0,phh), s = q/(2r), ), s = q/(2r), z > 0

distribution has density function qp + r

p is the source (target) value

s is the quadratic coefficient

z is the scale factor

DecodeDecode

z z z z22 + 4zsf + 4zsf

2zs2zs

ff [0, [0, f fhh), s = q/(2r), z > ), s = q/(2r), z > 00

f is the flattened value

s is the quadratic coefficient

z is the scale factor

Decode (f) =Decode (f) =

Order Preserving Order Preserving EncryptionEncryption

CipherteCiphertextxt

PlaintextPlaintext

11 2800028000

22 3500035000

…… ……

CCnn PPnn

Compute distinct Compute distinct attribute values in attribute values in ascending orderascending order

Ciphertext is the Ciphertext is the index valueindex value

• Effectively hides the Effectively hides the distribution of plaintext distribution of plaintext valuesvalues

•The key size is The key size is proportional to the proportional to the number of distinct number of distinct attribute valuesattribute values

•Any updates require Any updates require recomputing the key and recomputing the key and ciphertext valuesciphertext values

NoNo NameName PositioPositionn

SalarySalary LocatioLocationn

…… …… …… …… ……

Target Distribution Target Distribution RequirementRequirement Why isn’t the source-to-uniform Why isn’t the source-to-uniform

transformation sufficient for order transformation sufficient for order preserving encryption?preserving encryption?

It is, butIt is, but– The target distribution may cause an The target distribution may cause an

adversary to make incorrect assumptions adversary to make incorrect assumptions about the source distributionabout the source distribution

– The organization of the source distribution The organization of the source distribution cannot be inferred from the targetcannot be inferred from the target

Quadratic CoefficientQuadratic Coefficient

bb11 bb22

xx xx xx xx xx xx xx xx……

vvb1b1 – v – vb2b2

v v ==

xx xx

ii11 jj11 ii22 jj22

jj22 – i – i22 jj11 – i – i11vvj2j2 – v – vi2i2 vvj1j1 – v – vi1i1

--

q =q = s s = = 2 2 jj11 – i – i11

vvj1j1 – v – vi1i1

qq

Scale Factor Scale Factor ConstraintsConstraints

for all p [0,w) : M(p+1) – M(p) 2

wf = Kn

Ensures that there is a distinct mapped value for each input value

The width of a bucket in the mapped space is a function of the number of elements n in the bucket

K is the minimum width needed across buckets

Scale FactorScale Factor

Kn

sw2 + wz =

K = max [x(swi2+w)], i = 1, …, m,

x = 2, s 0

2/(1 + s(2w – 1)), s < 0

The scale factor will stretch short buckets to the width of the largest

bucket, further increasing the dimension of a bucket by a factor of the number of

elements in the bucket

SlopeSlope

b-1 b

The values within a single bucket are unevenly distributed within the bucket

Order Preserving Encryption for Numeric Data Rakesh Agrawal Jerry Kiernan Ramakrishnan Srikant Yirong Xu IBM Almaden Research Center.

Documents