Privacy-Preserving Data Mining Shared analysis without shared data. 16 February 2011 Chris Clifton

Transcription

1 Prvacy-Preservng Data Mnng Shared analyss wthout shared data 16 February 2011 Chrs Clfton

2 Idea: Collaborate to learn (only) desred results Data owners and FDA partcpate n protocol Results same as f all data sent to FDA Protocol ensures data not dsclosed Solutons for many types of analyss Theoretcally possble for any (polynomally computable) functon Protecton beyond ndvdual prvacy Even controls dsclosure of whch data owner responsble for whch event

3 Assocaton Rule Mnng: Horzontal Parttonng Goal: Learn condtons unusually lkely to lead to adverse outcome Low creatnne clearance & >0.125mg dgoxn hgh ADE * Identfy all such rules assocaton rule mnng Problem: rules occurrng at only one partcpatng ste could be lablty ssue Why s nsurer X the only one wth a partcular rule? Polces lmtng coverage of (possbly more approprate) medcatons? Soluton: Reveal only rules, not source But learn rules based on combned data from all sources * do: /aph.1e642

4 Overvew of the Method (Kantarcıoğlu and Clfton TKDE 04) Fnd the unon of the locally large canddate temsets securely Any rule suffcently strong at one ste that t mght be true globally But don t reveal source (or even number of stes where the rule s sgnfcant) Compute statstcs to determne global ncdence of rules E.g. A&B&C Adverse Drug Event But don t reveal any ste-specfc statstcs

5 Computng Canddate Sets Key commutatve encrypton E 1 (E 2 (E 3 (ABC))) E 1 (E 2 (E 3 (ABD))) 1 ABC E 1 (EE 12 (E 21 3 (ABD)) (ABC))) E 3 (ABC) E 3 (ABD) ABC ABD E 2 (EE 32 (E E 13 2 (ABC))) (ABD) 2 ABD 3 ABC E 3 (EE 13 (E E 21 3 (ABD))) (ABC))

6 Whch Rules are Globally Frequent? Goal: Gven a rule that s sgnfcant at (at least) one ste, s t sgnfcant overall? n X.sup s * DB (1) 1 n (2) 1 n (3) 1 Checkng (1) s equvalent to checkng (3) X.sup ( X.sup n 1 s* s* DB DB ) 0

7 Computng Frequent: Is A&B&CA.D.E 5%? 1 ABC: *300 ABC=18 ABC: 19 R? DBSze=300 FDA ABC: YES! 2 ABC=9 DBSze=200 3 ABC=5 DBSze=100 R=17 ABC: *200 ABC: R+count-freq.*DBSze *100

8 Prvacy-Preservng Data Mnng: Successes Numerous machne learnng tasks solved for horzontally and vertcally parttoned data Decson tree learnng and use K-Nearest Neghbor Clusterng: K-Means, EM, General dstancebased approaches Outler / anomaly detecton Collaboratve Flterng Naïve Bayes, Bayes network structure And many more

9 Prvacy-Preservng Data Mnng Shared analyss wthout shared data 16 February 2011 Chrs Clfton

10 Idea: Collaborate to learn (only) desred results Data owners and FDA partcpate n protocol Results same as f all data sent to FDA Protocol ensures data not dsclosed Solutons for many types of analyss Theoretcally possble for any (polynomally computable) functon Protecton beyond ndvdual prvacy Even controls dsclosure of whch data owner responsble for whch event

11 Assocaton Rule Mnng: Horzontal Parttonng Goal: Learn condtons unusually lkely to lead to adverse outcome Low creatnne clearance & >0.125mg dgoxn hgh ADE * Identfy all such rules assocaton rule mnng Problem: rules occurrng at only one partcpatng ste could be lablty ssue Why s nsurer X the only one wth a partcular rule? Polces lmtng coverage of (possbly more approprate) medcatons? Soluton: Reveal only rules, not source But learn rules based on combned data from all sources * do: /aph.1e642

12 Overvew of the Method (Kantarcıoğlu and Clfton TKDE 04) Fnd the unon of the locally large canddate temsets securely Any rule suffcently strong at one ste that t mght be true globally But don t reveal source (or even number of stes where the rule s sgnfcant) Compute statstcs to determne global ncdence of rules E.g. A&B&C Adverse Drug Event But don t reveal any ste-specfc statstcs

13 Computng Canddate Sets Key commutatve encrypton E 1 (E 2 (E 3 (ABC))) E 1 (E 2 (E 3 (ABD))) 1 ABC E 1 (EE 12 (E 21 3 (ABD)) (ABC))) E 3 (ABC) E 3 (ABD) ABC ABD E 2 (EE 32 (E E 13 2 (ABC))) (ABD) 2 ABD 3 ABC E 3 (EE 13 (E E 21 3 (ABD))) (ABC))

14 Whch Rules are Globally Frequent? Goal: Gven a rule that s sgnfcant at (at least) one ste, s t sgnfcant overall? n X.sup s * DB (1) 1 n (2) 1 n (3) 1 Checkng (1) s equvalent to checkng (3) X.sup ( X.sup n 1 s* s* DB DB ) 0

15 Computng Frequent: Is A&B&CA.D.E 5%? 1 ABC: *300 ABC=18 ABC: 19 R? DBSze=300 FDA ABC: YES! 2 ABC=9 DBSze=200 3 ABC=5 DBSze=100 R=17 ABC: *200 ABC: R+count-freq.*DBSze *100

16 Prvacy-Preservng Data Mnng: Successes Numerous machne learnng tasks solved for horzontally and vertcally parttoned data Decson tree learnng and use K-Nearest Neghbor Clusterng: K-Means, EM, General dstancebased approaches Outler / anomaly detecton Collaboratve Flterng Naïve Bayes, Bayes network structure And many more