1. Privacy Preserving Mining of Association Rules Alexandre Evfimievski, Ramakrishnan Srikant, Rakesh Agrawal, Johannes Gehrke IBM Almaden Research Center Cornell University

2. Data Mining and Privacy The primary task in data mining: development of models about aggregated data. Can we develop accurate models without access to precise information in individual data records?

3. Data Mining and Privacy The primary task in data mining: development of models about aggregated data. Can we develop accurate models without access to precise information in individual data records? Answer: yes, by randomization. –R. Agrawal, R. Srikant “Privacy Preserving Data Mining,” SIGMOD 2000 –for numerical attributes, classification How about association rules?

4. Randomization Overview Recommendation Service Alice Bob B. Spears, baseball, cnn.com, … J.S. Bach, painting, nasa.gov, … Chris B. Marley, camping, linux.org, …

5. Recommendation Service Alice Bob J.S. Bach, painting, nasa.gov, … J.S. Bach, painting, nasa.gov, … B. Spears, baseball, cnn.com, … B. Spears, baseball, cnn.com, … B. Marley, camping, linux.org, … B. Marley, camping, linux.org, … B. Spears, baseball, cnn.com, … J.S. Bach, painting, nasa.gov, … Chris B. Marley, camping, linux.org, … Randomization Overview

6. Recommendation Service Associations Recommendations Alice Bob J.S. Bach, painting, nasa.gov, … J.S. Bach, painting, nasa.gov, … B. Spears, baseball, cnn.com, … B. Spears, baseball, cnn.com, … B. Marley, camping, linux.org, … B. Marley, camping, linux.org, … B. Spears, baseball, cnn.com, … J.S. Bach, painting, nasa.gov, … Chris B. Marley, camping, linux.org, … Randomization Overview

7. Recommendation Service Associations Recommendations Alice Bob Metallica, painting, nasa.gov, … Metallica, painting, nasa.gov, … B. Spears, soccer, bbc.co.uk, … B. Spears, soccer, bbc.co.uk, … B. Marley, camping, microsoft.com … B. Marley, camping, microsoft.com … B. Spears, baseball, cnn.com, … J.S. Bach, painting, nasa.gov, … Support Recovery Chris B. Marley, camping, linux.org, … Randomization Overview

8. Associations Recap A transaction t is a set of items (e.g. books) All transactions form a set T of transactions Any itemset A has support s in T if Itemset A is frequent if s  s min If A  B, then supp (A)  supp (B).

9. Associations Recap A transaction t is a set of items (e.g. books) All transactions form a set T of transactions Any itemset A has support s in T if Itemset A is frequent if s  s min If A  B, then supp (A)  supp (B). Example: –20% transactions contain beer, –5% transactions contain beer and diapers; –Then: confidence of “beer  diapers” is 5/20 = 0.25 = 25%.

10. The Problem How to randomize transactions so that –we can find frequent itemsets –while preserving privacy at transaction level?

11. Talk Outline Introduction Privacy Breaches Our Solution Experiments Conclusion

12. Uniform Randomization Given a transaction, –keep item with 20% probability, –replace with a new random item with 80% probability.

13. Example: {x, y, z} 1% have {x, y, z} 5% have {x, y}, {x, z}, or {y, z} only 10 M transactions of size 10 with 10 K items: 94% have one or zero items of {x, y, z}

14. Example: {x, y, z} 1% have {x, y, z} 5% have {x, y}, {x, z}, or {y, z} only 10 M transactions of size 10 with 10 K items: 94% have one or zero items of {x, y, z} Uniform randomization: How many have {x, y, z} ?

15. Example: {x, y, z} 1% have {x, y, z} 5% have {x, y}, {x, z}, or {y, z} only 10 M transactions of size 10 with 10 K items: 94% have one or zero items of {x, y, z} 0.008% 800 ts. 0.00016% 16 trans. less than 0.00002% 2 transactions Uniform randomization: How many have {x, y, z} ? 0.2 2 8/10,000 0.2 3 at most 0.2 (9/10,000) 2

16. Example: {x, y, z} 1% have {x, y, z} 5% have {x, y}, {x, z}, or {y, z} only 10 M transactions of size 10 with 10 K items: 94% have one or zero items of {x, y, z} 0.008% 800 ts. 97.8% 0.00016% 16 trans. 1.9% less than 0.00002% 2 transactions 0.3% Uniform randomization: How many have {x, y, z} ? 0.2 2 8/10,000 0.2 3 at most 0.2 (9/10,000) 2

17. Example: {x, y, z} Given nothing, we have only 1% probability that {x, y, z} occurs in the original transaction Given {x, y, z} in the randomized transaction, we have about 98% certainty of {x, y, z} in the original one. This is what we call a privacy breach. Uniform randomization preserves privacy “on average,” but not “in the worst case.”

18. Privacy Breaches Suppose: –t is an original transaction; –t’ is the corresponding randomized transaction; –A is a (frequent) itemset. Definition: Itemset A causes a privacy breach of level  (e.g. 50%) if, for some item z  A, –Assumption: no external information besides t’.

19. Talk Outline Introduction Privacy Breaches Our Solution Experiments Conclusion

20. Our Solution Insert many false items into each transaction Hide true itemsets among false ones “Where does a wise man hide a leaf? In the forest. But what does he do if there is no forest?” “He grows a forest to hide it in.” G.K. Chesterton

21. Our Solution Insert many false items into each transaction Hide true itemsets among false ones Can we still find frequent itemsets while having sufficient privacy? “Where does a wise man hide a leaf? In the forest. But what does he do if there is no forest?” “He grows a forest to hide it in.” G.K. Chesterton

22. Definition of cut-and-paste Given transaction t of size m, construct t’ : a, b, c, u, v, w, x, y, z t = t’ =

23. Definition of cut-and-paste Given transaction t of size m, construct t’ : –Choose a number j between 0 and K m (cutoff); a, b, c, u, v, w, x, y, z t = t’ = j = 4

24. Definition of cut-and-paste Given transaction t of size m, construct t’ : –Choose a number j between 0 and K m (cutoff); –Include j items of t into t’ ; a, b, c, u, v, w, x, y, z t = b, v, x, z t’ = j = 4

25. Definition of cut-and-paste Given transaction t of size m, construct t’ : –Choose a number j between 0 and K m (cutoff); –Include j items of t into t’ ; –Each other item is included into t’ with probability p m. The choice of K m and p m is based on the desired level of privacy. a, b, c, u, v, w, x, y, z t = b, v, x, z t’ = œ, å, ß, ξ, ψ, €, א, ъ, ђ, … j = 4

26. Partial Supports To recover original support of an itemset, we need randomized supports of its subsets. Given an itemset A of size k and transaction size m, A vector of partial supports of A is –Here s k is the same as the support of A. –Randomized partial supports are denoted by

27. Transition Matrix Let k = |A|, m = |t|. Transition matrix P = P (k, m) connects randomized partial supports with original ones: Randomized supports are distributed as a sum of multinomial distributions.

28. The Unbiased Estimators Given randomized partial supports, we can estimate original partial supports:

29. The Unbiased Estimators Given randomized partial supports, we can estimate original partial supports: Covariance matrix for this estimator: To estimate it, substitute s l with (s est ) l. –Special case: estimators for support and its variance

30. Class of Randomizations Our analysis works for any randomization that satisfies two properties: –A per-transaction randomization applies the same procedure to each transaction, using no information about other transactions; –An item-invariant randomization does not depend on any ordering or naming of items.

31. Class of Randomizations Our analysis works for any randomization that satisfies two properties: –A per-transaction randomization applies the same procedure to each transaction, using no information about other transactions; –An item-invariant randomization does not depend on any ordering or naming of items. Both uniform and cut-and-paste randomizations satisfy these two properties.

32. Apriori Let k = 1, candidate sets = all 1-itemsets. Repeat: 1.Count support for all candidate sets 2.Output the candidate sets with support  s min 3.New candidate sets = all (k + 1) -itemsets s.t. all their k -subsets are candidate sets with support  s min 4.Let k = k + 1 Stop when there are no more candidate sets.

33. The Modified Apriori Let k = 1, candidate sets = all 1-itemsets. Repeat: 1.Estimate support and variance (σ 2 ) for all candidate sets 2.Output the candidate sets with support  s min 3.New candidate sets = all (k + 1) -itemsets s.t. all their k -subsets are candidate sets with support  s min - σ 4.Let k = k + 1 Stop when there are no more candidate sets, or the estimator’s precision becomes unsatisfactory.

34. Privacy Breach Analysis How many added items are enough to protect privacy? –Have to satisfy Pr [z  t | A  t’] <  (  no privacy breaches) –Select parameters so that it holds for all itemsets. –Use formula ( ):

35. Privacy Breach Analysis How many added items are enough to protect privacy? –Have to satisfy Pr [z  t | A  t’] <  (  no privacy breaches) –Select parameters so that it holds for all itemsets. –Use formula ( ): Parameters are to be selected in advance! –Construct a privacy-challenging test: an itemset whose all subsets have maximum possible support. –Enough to know maximal support of an itemset for each size.

36. Graceful Tradeoff Want more precision or more privacy? –Adjust privacy breach level –A small relaxation of privacy restrictions results in a small increase in precision of estimators.

37. Talk Outline Introduction Privacy Breaches Our Solution Experiments –Support recovery vs. parameters –Real-life data Conclusion

38. Lowest Discoverable Support LDS is s.t., when predicted, is 4  away from zero. Roughly, LDS is proportional to |t| = 5,  = 50%

39. LDS vs. Breach Level |t| = 5, |T| = 5 M Reminder: breach level is the limit on Pr [z  t | A  t’]

40. Talk Outline Introduction Privacy Breaches Our Solution Experiments –Support recovery vs. parameters –Real-life data Conclusion

41. Real datasets: soccer, mailorder Soccer is the clickstream log of WorldCup’98 web site, split into sessions of HTML requests. –11 K items (HTMLs), 6.5 M transactions –Available at http://www.acm.org/sigcomm/ITA/http://www.acm.org/sigcomm/ITA/ Mailorder is a purchase dataset from a certain on-line store –Products are replaced with their categories –96 items (categories), 2.9 M transactions A small fraction of transactions are discarded as too long. –longer than 10 (for soccer) or 7 (for mailorder)

42. Modified Apriori on Real Data Itemset Size True Itemsets True Positives False Drops False Positives 12662541231 22171952245 34843526 Itemset Size True Itemsets True Positives False Drops False Positives 165 00 22282121628 3221845 Soccer: s min = 0.2%   0.07% for 3-itemsets Mailorder: s min = 0.2%   0.05% for 3-itemsets Breach level = 50%. Inserted 20-50% items to each transaction.

43. False Drops False Positives Size< 0.10.1-0.150.15-0.2  0.2 1 0210254 2 0517195 3 01443 Size< 0.10.1-0.150.15-0.2  0.2 1 0724254 2 71028195 3 513843 Size< 0.10.1-0.150.15-0.2  0.2 1 00065 2 0115212 3 01318 Size< 0.10.1-0.150.15-0.2  0.2 1 00065 2 0028212 3 12218 Soccer Mailorder Pred. supp%, when true supp  0.2% True supp%, when pred. supp  0.2%

44. Actual Privacy Breaches Verified actual privacy breach levels The breach probabilities are counted in the datasets for frequent and near-frequent itemsets. If maximum supports were estimated correctly, even worst-case breach levels fluctuated around 50% –At most 53.2% for soccer, –At most 55.4% for mailorder.

45. Talk Outline Introduction Privacy Breaches Our Solution Experiments Conclusion

46. Summary Privacy breaches: identified problem and provided a solution for controlling breaches Derived estimators of support and variance for a class of randomization operators Algorithm for discovering associations in randomized data Validated on real-life datasets Can find associations while preserving privacy at the level of individual transactions

47. Future Work Control of more general privacy breaches –What about other properties of transactions, for example item z  t breach caused by A  t’ =  ? –What about external information? Theoretical limits of discoverability for a given privacy breach level –How to compute theoretical limits? –How to attain them by an algorithm?

48. Thank You!

49. BACK-UPS

50. Our Solution: Example Old set-up: –Given 10,000 items, 10 M transactions of size 10 –100,000 transactions (1%) contain A = {x, y, z} In addition to uniform randomization with p = 80%, insert 500 new random items to each transaction. –~ 800 transactions contain {x, y, z} before and after; –Roughly (10 M) (500 / 10,000) 3 = 1250 transactions contain none before and full {x, y, z} after. Presence of {x, y, z} in a randomized transaction now says little about the original transaction.

51. Privacy Breach Analysis GIVEN: itemset A, and item z  A WANTED: Assume that partial supports are probabilities: Define: Then we have:

52. Limiting Privacy Breaches We want to make sure that always But we do not know supports in advance. Solution: For each itemset size k, give “privacy- challenging” test values to. –It is an itemset whose subsets have maximum supports –We need to estimate maximum support values prior to randomization

53. LDS vs. Transaction Size  = 50%, |T| = 5 M Too long transactions cannot be used for prediction

54. Related Work R. Agrawal, R. Srikant “Privacy Preserving Data Mining,” SIGMOD 2000: Each client has a numerical attribute x i Client i sends x i + y i, where y i = random offset, with known distribution Server reconstructs the distribution of original attributes (~ EM algorithm) The distribution is then used for classification –Numerical attributes only

55. Related Work Y. Lindell and B. Pinkas “Privacy Preserving Data Mining,” Crypto 2000 J. Vaidya and C. Clifton “Privacy Preserving Association Rule Mining in Vertically Partitioned Data” …

56. Privacy Concern Popular press: –“The End of Privacy”, “The Death of Privacy” Government directives: –European directive on privacy protection (Oct 98) –Canadian Personal Information Protection Act (Jan 2001) Surveys of Web users: –17% fundamentalists, 56% pragmatic majority, 27% marginally concerned (April 99) –82% said having privacy would matter (July 99)