K.U.Leuven George Danezis 1, Markulf Kohlweiss 1, Ben Livshits 1, and Alfredo Rial 2 Private Client-Side Profiling with Random Forests and Hidden Markov.

K.U.Leuven

George Danezis1, Markulf Kohlweiss1, Ben Livshits1, and Alfredo Rial2

Private Client-Side Profiling with Random Forests and Hidden Markov Models

1Microsoft Research2KU Leuven ESAT/COSIC – IBBT, Belgium

PETS 2012

Private Client-Side Profiling PETS 2012

PETS 2012

• Introduction• System Overview• Applications• Random Forests• Our Protocol• Conclusion

Index

2Private Client-Side Profiling

http://www.dmrdirect.com/direct-mail/customer-profiling/gain-valuable-marketing-intelligence/

PETS 2012

1 – Introduction

3Private Client-Side Profiling

http://blog.maia-intelligence.com/2009/10/05/customer-analytics-in-retail/

PETS 2012

• Client Profiling -> Deliver Customized Services• Current techniques:

o Cookieso Third party apps in social networkso Web bugs

• Disadvantageso Privacyo Correctness

• Ad-hoc• Block

Current Client Profiling Tools

4Private Client-Side Profiling

http://www.pc-xp.com/2010/12/04/web-bug-reveals-internet-browsing-history/

PETS 2012

• User’s perform the classification task:o Input certified features and certified algorithmo Run algorithm:

• Classification: Random Forest• Pattern Recognition: Hidden Markov Model

o Output result and proof of correctnesso Service provider verifies result

• Advantageso Privacy: Only classification result is disclosedo Correctness guaranteed by proof

Private Client-Side Profiling

5Private Client-Side Profiling

PETS 2012

2- System Overview

6Private Client-Side Profiling

PETS 2012

• Behavioral advertising• P2P dating & matchmaking• Financial logs• Pay-as-you-drive Insurance• Bio-medical & genetic

3- Applications

7Private Client-Side Profiling

PETS 2012

Behavioural Advertising

8Private Client-Side Profiling

http://kickstand.typepad.com/metamuse/2008/05/behavioral-adve.html

PETS 2012

P2P Dating & Matchmaking

9Private Client-Side Profiling

http://www.robhelsby.com/P2P%20Dating.html

PETS 2012

Financial logs

10Private Client-Side Profiling

http://www.ikeepsafe.org/privacy/arm-yourself-against-online-fraud/

PETS 2012

Pay-as-you-drive Insurance

11Private Client-Side Profiling

http://www.fenderbender.com/FenderBender/April-2011/Pay-As-You-Drive-Insurance/

PETS 2012

Bio-medical & Genetic

12Private Client-Side Profiling

http://www.pattern-expert.com/Bioinformatics/eng/bioinformatics/SNPAnalysis.html

PETS 2012

4- Random Forests

13Private Client-Side Profiling

http://www.iis.ee.ic.ac.uk/~tkkim/iccv09_tutorial.html

PETS 2012

• Classification algorithm: a data item with a set of features is classified into two classes or

• It consists of a collection of trees. Each tree:oNon-leaf nodes: oLeaf-nodes:

• Classification result:

Definition of Random Forest

14Private Client-Side Profiling

PETS 2012

Tree Example

15Private Client-Side Profiling

PETS 2012

• Zero-Knowledge Proofs of Knowledge

• P-Signatures: signature schemes with an efficient ZKPK of signature possession

5- Our Protocol

16Private Client-Side Profiling

PETS 2012

• LOOKUP

• ZKTABLE

Notation

17Private Client-Side Profiling

PETS 2012

• A sends Prover his certified features:

Phase 1

18Private Client-Side Profiling

PETS 2012

A’ sends Prover a certified random forest:• Branches:

o Left Branches:o Right Branches:

• Leaf nodes:

Phase 2

19Private Client-Side Profiling

PETS 2012

• Prover computes the following ZKPK:

Phase 3 – Tree Resolution

20Private Client-Side Profiling

PETS 2012

• Prover repeats tree resolution for all the trees

Phase 3 – Forest Resolution

21Private Client-Side Profiling

PETS 2012

• P-signature scheme by Au et al. [SCN 2006]• Hidden range proof based on Camenisch et al.

[Asiacrypt 2008]• Random forest parameters:oNumber of trees: t = 50oDepth: D = 10oNumber of features: M = 100oAverage number of feature values: K = 100

Instantiation

22Private Client-Side Profiling

PETS 2012

• Fu = Table of certified user features

• Bt = Table of branches of all trees

• Lt = Table of leaf nodes of all trees

• Vt = Table of signatures for the hidden range proof

• Pt = Proof of random forest resolution

Efficiency

23Private Client-Side Profiling

PETS 2012

• Private Client-Side Profiling:o Classification: Random Forestso Pattern Recognition: Hidden Markov Models

• The mere act of profiling may violate privacy.

Conclusion

24Private Client-Side Profiling

“We do not see the power which is in speech because we forget that all speech is a classification, and thatAll classifications are oppressive”

Roland Barthes

PETS 2012

Comparison Shopping

25Private Client-Side Profiling

http://article.wn.com/view/2012/04/19/Life_insurance_cos_new_biz_premiums_down_92/