1. Secure Efficient Multiparty Computing of Multivariate Polynomials and Applications Dana Dachman-Soled, Tal Malkin, Mariana Raykova, Moti Yung

2. 2 x1x1 x2x2 x3x3 x4x4

3. 3 x1x1 x2x2 x3x3 x4x4 F 1 (x 1,x 3,x 3 ) F 2 (x 1,x 3,x 3 ) F 3 (x 1,x 3,x 3 ) F 4 (x 1,x 3,x 3 )

4. 4 Secure Multiparty Computation How to compute a function on the private inputs of multiple parties not leaking more than the result? Secure Multiparty Computation How to compute a function on the private inputs of multiple parties not leaking more than the result?

5. 5 Secure Multiparty Computation Feasible – [Yao82], [GMW87], [CDv88], [BG89], [BG90], [Cha90], [Bea92], … Not Efficient – communication and computation proportional to circuit size Secure Multiparty Computation Feasible – [Yao82], [GMW87], [CDv88], [BG89], [BG90], [Cha90], [Bea92], … Not Efficient – communication and computation proportional to circuit size

6. 6 x1x1 x2x2 x3x3 x4x4 Multivariate Polynomials

7. 7 x1x1 x2x2 x3x3 x4x4 Applications

8. 8 x1x1 x2x2 x3x3 x4x4 Multivariate Polynomials Applications Multiparty Set Intersection

9. 9 x1x1 x2x2 x3x3 x4x4 Multivariate Polynomials Applications Linear Algebra matrix arithmetic, inverse, determinant, Eigen values

10. 10 x1x1 x2x2 x3x3 x4x4 Multivariate Polynomials Applications Statistics functions average, standard deviation, variance, chi-square test, computing Pearson’s correlation coefficients

11. 11 x1x1 x2x2 x3x3 x4x4 Multivariate Polynomials Applications Taylor series approximation trigonometric functions, logarithms, exponents, square root

12. 12 Outsourced computation many workers at least one honest

13. 13 Outsourced computation Computation on shares, Reconstruction of output

14. Our results Multiparty computation protocol for functionalities that can be represented as multivariate polynomials – Improvement of generic complexity for multiple parties Left as open problem in FM10 Security: – Against malicious majority – Proofs in the standard simulation model Black box construction from homomorphic encryption with a natural property…. – Instantiated through threshold Paillier encryption (decisional composite residuosity) 14

15. Our Results Efficiency: – Communication complexity – FM10 subexponential in the number of parties, we achieve fully polynomial (in all parameters) complexity: Broadcast complexity Round table complexity – Constant number round table rounds Application construction: Multiparty Set Intersection – Improve complexity of existing multiparty solutions KS05, SS09, CJS10 15

16. Building Blocks Input sharing using committed Shamir/Reed- Solomon codes P X (0) = X shares P X (1), …, P X (D) Vector Homomorphic Encryption ENC(m 1 ; r 1 ) ⊗ ENC(m 2 ; r 2 ) = ENC(m 1 + m 2 ; r 1 ⊕ r 2 ) ENC(m; r) c = ENC(c · m; r ⊙ c) – Instantiation: threshold Paillier encryption 16

17. Building Blocks Polynomial code commutativity Interpolate (Poly-Eval (inputs shares)) = Poly-Eval (Interpolate (inputs shares)) = Poly-Eval(inputs) Incremental encrypted polynomial evaluation – Each monomial M = c  i=1 h i (inputs of party i) – b 0 = ; = ⊕ 17 b i+1 Enc(c) bibi bibi h i (inputs of party i) #parties Encryption of partial evaluation of M with inputs from first i+1/i parties Constant for homomorphic property

18. Building blocks Lagrange Interpolation Protocol Over Encrypted Values: – given A > d+1 encrypted points (1, ENC pk (y 1, r 1 )),... (A, ENC pk (y A, r A )) – check that they lie on poly of degree d ENC pk (y i,r i ) =  j=1 (ENC pk (y j,r j )) L j (i) – synchronized randomness Randomness Interpolation – given (1,y 1 ),...,(A,y A ),r 1,...,r d+1 – compute r d+2,..., r A – Encrypted interpolation holds for [i, ENC pk (y i, r i )] 1≤i≤A d+1 18

19. Efficient Input Preprocessing Polynomial Degree Reduction Change of variables Polynomial Q(y) of degree n Q(y) Q(y 0,y 1,y 2 …, y  log n  ) y 0 = y y 1 = y 2 y 2 = y 4 ………. y  log n  = y 2  log n  Deg: nDeg: log n y 19

20. Proof of Knowledge and Verification Correct computation of new variables Correct degree of input sharing polynomials Prover: x 1,…,x n Common: c 1,…,c n, L (x 1,…,x n )  L c i = ENC(x i ) InputProof Output Verifier: Accept/Reject enc(r 1 ) enc(r 2 ) enc(r n ) c 1 * enc(r 1 ) c 2 * enc(r 2 ) … c n * enc(r n ) (x 1 +r 1,…,x n +r n )  L (r 1,…,r n )  L open 0 1 … c i * enc(r i ) = enc(x i +r i ) 20

21. Protocol Outline 21

22. Efficient preprocessing for each variable in the multivariate polynomial Commit to shares of new variables 22

23. Each party P i contributes his inputs – in each monomial s for each share j = · 23 b i+1,j,s b i,j,s ⊕ h i (share j of P i ) Enc(0, r i,j,s ) r i,j,s generated with randomness interpolation protocol

24. Each party re-randomizes the final output shares S 1, …, S 10kD – Randomizng polynomial P j,0 (0) = 0 – Shares (1,P j,0 (1)),...,(10kD,P j,0 (10kD)) – Re-randomized output shares = · 24 S’i S’i S’i S’i Si Si Si Si  j=1 ENC pk (P j,0 (i);r j,i ) m r j,kD+2,...,r j,10kD generated with randomness interpolation protocol

25. All parties verify that the encrypted output shares S i lie on a polynomial of degree kD Parties select a subset of the shares of size k and decommit corresponding shares Parties verify the computation of the open shares 25 P 1 (1) P 2 (1) Com(P 1 (2)) Com(P 2 (2)) Com(P 1 (3)) Com(P 2 (3)) P 1 (1) P 2 (4) Com(P 1 (10kD) ) Com(P 2 (10kD) ) … … Verify computation Verify degree

26. The parties run threshold decryption for each of the output shares The output receiver interpolates the output value from the shares 26

27. Protocol Complexities Amortized – sharing with multiple secrets Communication complexity – Round table – between consecutive parties: intermediate protocol messages O(Dn(m-1)), m parties, n monomials, D sum of log variable degrees – Broadcast – input commitments, decommitments in verification phase Smaller than polynomial representation O(D (  j=1  j=1 log α j,t )) α j,t highest degree of variable, L j inputs for party j Computational complexity O(Dnm) mLjLj 27

28. Multiparty set intersection = · + Optimizations: – Only two parties have inputs per each monomial – Inputs that are used only once do not need to be shared Complexity - m parties, d inputs each: – Communication - O(md + 10d log 2 d); CJS10 – quadratic in number of parties, other solutions worse complexity – Computation - O(md 2 log d) 28 P(x) ri ri ri ri P i (x) x x r i = r i,1 + … + r i,m r i,j randomness from party j P i (x) represents the input set of party i  j=1 m-1

29. Thank You! Questions? 29