1. P4P: A Practical Framework for Privacy- Preserving Distributed Computation Yitao Duan (Advisor Prof. John Canny) http://www.cs.berkeley.edu/~duan Berkeley Institute of Design Computer Science Division University of California, Berkeley 11/27/2007

2. Research Goal To provide practical solutions with provable privacy and adequate efficiency in a realistic adversary model at reasonably large scale

3. Research Goal To provide practical solutions with provable privacy and adequate efficiency in a realistic adversary model at reasonably large scale

4. Model …… u1u1 d1d1 u2u2 d2d2 u n-1 d n-1 unun dndn f Challenge: standard cryptographic tools not feasible at large scale Must be obfuscated

5. A Practical Solution Provable privacy: Cryptography Efficiency: Minimize the number of expensive primitives and rely on probabilistic guarantee Realistic adversary model: Must handle malicious users who may try to bias the computation by inputting invalid data

6. Basic Approach …… u1u1 d1d1 u2u2 d2d2 u n-1 d n-1 unun dndn Σ g j (d i ) g j (d n ) g j (d 2 )g j (d n-1 ) Cryptographic privacy f = d i in D, g j : j = 1, 2, …, m No leakage beyond final result for many algorithms

7. The Power of Addition A large number of popular algorithms can be run with addition-only steps Linear algorithms: voting and summation, nonlinear algorithm: regression, classification, SVD, PCA, k-means, ID3, EM etc All algorithms in the statistical query model [Kearns 93] Many other gradient-based numerical algorithms Addition-only framework has very efficient private implementation in cryptography and admits efficient ZKPs

8. Peers for Privacy: The Nomenclature Privacy is a right that one must fight for. Some agents must act on behalf of user’s privacy in the computation. We call them privacy peers Our method aggregates across many user data. We can prove that the aggregation provides privacy: the data from the peers protects each other

9. Private Addition – P4P Style The computation: secret sharing over small field Malicious users: efficient zero-knowledge proof to bound the L2-norm of the user vector

10. Big Integers vs. Small Ones Most applications work with “regular-sized” integers (e.g. 32- or 64-bit). Arithmetic operations are very fast when each operand fits into a single memory cell (~10 -9 sec) Public-key operations (e.g. used in encryption and verification) must use keys with sufficient length (e.g. 1024-bit) for security. Existing private computation solutions must work with large integers extensively (~10 -3 sec) A 6 orders of magnitude difference!

11. Private Arithmetic: Two Paradigms Homomorphism: User data is encrypted with a public key cryptosystem. Arithmetic on this data mirrors arithmetic on the original data, but the server cannot decrypt partial results. Secret-sharing: User sends shares of their data to several servers, so that no small group of servers gains any information about it.

12. Arithmetic: Homomorphism vs VSS Homomorphism + Can tolerate t < n corrupted players as far as privacy is concerned - Use public key crypto, works with large fields (e.g. 1024-bit), 10,000x more expensive than normal arithmetic (even for addition) Secret sharing + Addition is essentially free. Can use any size field - Can’t do two party multiplication - Most schemes also use public key crypto for verification - Doesn’t fit well into existing service architecture

13. P4P: Peers for Privacy Some parties, called Privacy Peers, actively participate in the computation, working for users’ privacy Privacy peers provide privacy when they are available, but cant access data themselves P P U U U U U U S Peer Group

14. P4P The server provides data archival, and synchronizes the protocol Server only communicates with privacy peers occasionally (2AM) P P U U U U U U S Peer Group

15. Privacy Peers Roles of privacy peers Anonymizing communication Sharing information Participating in computation Others infrastructure support They work on behalf of users privacy But we need a higher level of trust on privacy peers

16. Candidates for Privacy Peers Some players are more trustworthy than others In workspace, a union representative In a community, a few members with good reputation Or a third party commercial provider A very important source of security and efficiency The key is that privacy peers should have different incentives from the server, a mutual distrust between them

17. Security from Heterogeneity Server is secure against outside attacks and won’t actively cheat Companies spend $$$ to protect their servers The server often holds much more valuable info than what the protocol reveals Server benefits from accurate computation Privacy peers won’t collude with the server Interests conflicts, mutual distrust, laws Server can’t trust clients can keep conspiracy secret Users can actively cheat Rely on server for protection against outside attacks, privacy peers for defending against a curious server

18. vivi uiui u i + v i = d i Private Addition d i : user i’s private vector. u i,,v i and d i are all in a small integer field

19. μ = Σu i ν = Σv i u i + v i = d i Private Addition

20. u i + v i = d i μ = Σu i ν = Σv i μ ν Private Addition

21. μ + ν Private Addition

22. Provable privacy Computation on both the server and the privacy peer is over small field: same cost as non-private implementation Fits existing server-based schemes Server is always online. Users and privacy peers can be on and off. Only two parties performing the computation, users just submit their data (and provide a ZK proof, see later) Extra communication for the server is only with the privacy peer, independent of n P4P’s Private Addition

23. The Need for Verification This scheme has a glaring weakness. Users can use any number in the small field as their data. Think of a voting scheme: “Please place your vote 0 or 1 in the envelope” Bush 100,000 Gore -100,000

24. Zero Knowledge Proofs I can prove that I know X without disclosing what X is. I can prove that a given encrypted number is a 0. Or I can prove that an encrypted number is a 1. I can prove that an encrypted number is a ZERO OR ONE, i.e. a bit. (6 extra numbers needed) I can prove that an encrypted number is a k-bit integer. I need 6k extra numbers to do this (!!!)

25. An Efficient ZKP of Boundedness Luckily, we don’t need to prove that every number in a user’s vector is small, only that the vector is small. The server asks for some random projections of the user’s vector, and expects the user to prove that the square sum of them is small. O(log m) public key crypto operations (instead of O(m)) to prove that the L-2 norm of an m-dim vector is smaller than L. Running time reduced from hours to seconds.

26. Bounding the L2-Norm A natural and effective way to restrict a cheating user’s malicious influence You must have a big vector to produce large influence on the sum Perturbation theory bounds system change with norms: |σ i (A) - σ i (B)| ≤ ||A-B|| 2 [ Weyl] Can be the basis for other checks Setting L = 1 forces each user to have only 1 vote

27. Random Projection-based L2-Norm ZKP Server generates N random m -vectors in {-1, 0, +1} m User projects his data to the N directions. provides ZKP that the square sum of the projections < NL 2 /2 Expensive public key operations are only on the projections and the square sum

28. Effectiveness

29. Acceptance/rejection probabilities (a) Linear and (b) log plots of probability of user input acceptance as a function of |d|/L for N = 50. (b) also includes probability of rejection. In each case, the steepest (jagged curve) is the single-value vector (case 3), the middle curve is Zipf vector (case 2) and the shallow curve is uniform vector (case 1)

30. Performance Evaluation (a) Verifier and (b) prover times in seconds for the validation protocol where (from top to bottom) L (the required bound) has 40, 20, or 10 bits. The x-axis is the vector length.

31. SVD Singular value decomposition is an extremely useful tool for a lot of IR and data mining tasks (CF, clustering … ) SVD for a matrix A is a factorization A = UDV T. If A encodes users x items, then V T gives us the best least-squares approximations to the rows of A in a user-independent way. A T AV = VD  SVD is an eigenproblem

32. SVD: P4P Style

33. Experiments: SVD Datasets DatasetDimensionsDensityRangeType Enron150×1500.0736[0,1593]Social graph EM74424×16480.0229[0, 1.0]Movie ratings RAND2000×20001.0[-2 20, 2 20 ]Random

34. Results N : number of iterations. k : number of singular values. ε: relative residual error

35. Distributed Association Rule Mining n users, m items. User i has dataset D i Horizontally partitioned: D i contains the same attributes 1 0 0 0 1 …… 0 0 0 0 1 0 0 …… 1 0 …… D1D1 DnDn

36. The Market-Basket Model A large set of items, e.g., things sold in a supermarket. A large set of baskets, each of which is a small set of the items, e.g., the things one customer buys on one day.

37. Support Simplest question: find sets of items that appear “frequently” in the baskets. Support for itemset I = the number of baskets containing all items in I. Given a support threshold s, sets of items that appear in > s baskets are called frequent itemsets.

38. Example Items={milk, coke, pepsi, beer, juice}. Support = 3 baskets. B 1 = {m, c, b}B 2 = {m, p, j} B 3 = {m, b}B 4 = {c, j} B 5 = {m, p, b}B 6 = {m, c, b, j} B 7 = {c, b, j}B 8 = {b, c} Frequent itemsets: {m}, {c}, {b}, {j}, {m, b}, {c, b}, {j, c}.

39. Association Rules If-then rules about the contents of baskets. {i 1, i 2,…,i k } → j means: “if a basket contains all of i 1,…,i k then it is likely to contain j.” Confidence of this association rule is the probability of j given i 1,…,i k.

40. Step k of apriori-gen in P4P User i constructs an m k -Dimensional vector in small field ( m k : number of candidate itemset at step k ) Use P4P to compute the aggregate (with verification) The result encodes the supports of all candidate itemsets

41. Step k of apriori-gen in P4P 1 0 0 0 1 …… 0 0 0 0 1 0 0 …… 1 0 …… D1D1 DnDn c j : j th candidate itemset * + d1[j]d1[j] + dn[j]dn[j] P4P Support for c j

42. Analysis Privacy guaranteed by P4P Near optimal efficiency: cost comparable to that of a direct implementation of the algorithms Main aggregation in small field Only a small number of large field operations Deal with cheating users with P4P’s built- in ZK user data verification

43. Privacy SVD: The intermediate sums are implied by the final results A T A = VDV T ARM: Sums treated as public by the applications Guaranteed privacy regardless data distribution or size

44. Infrastructure Support Multicast encryption [RSA 06] Scalable secure bidirectional communication [Infocom 07] Data protection scheme [PET 04]

45. P4P: Current Status P4P has been implemented In Java using native code for big integer Runs on Linux platform Will be made an open-source toolkit for building privacy-preserving real-world applications.

46. Conclusion We can provide strong privacy protection with little or no cost to a service provider for a broad class of problems in e-commerce and knowledge work. Responsibility for privacy protection shifts to privacy peers Within the P4P framework, private computation and many zero-knowledge verifications can be done with great efficiency

47. More info duan@cs.berkeley.edu http://www.cs.berkeley.edu/~duan Thank You!