A Simple BGN-Type Cryptosystem from LWE Craig Gentry Shai Halevi Vinod Vaikuntanathan IBM Research Perspective Homomorphic Encryption in three easy steps [G’09] • Step 1: Encryption from linear codes • SK/PK are Good/Bad representation of code • Bad representation, can’t tell words close to code from random • Good representation can be used to correct many errors • Additive homomorphism “for free” • Step 2: ECC lives inside a ring • We have both additive, multiplicative sructure • If code is an ideal, also multiplicative homomorphism • for low-degree polynomials • Step 3: Bootstrapping, Squashing, etc. Instances of this Paradigm • Ring of polynomials [G’09] • Ring of integers [vDGHV’10] • This work: how about ring of matrices? • • • • • Doesn’t quite work like the others We only get additive-HE + one multiplication Quadratic formulas, as in [BGN’05] But more efficient and more flexible Can be made leakage-resilient, identity-based Background Learning with Errors (LWE) • n – security parameter • q poly(n) • m > n log q n s c =m A random mod q + x mod q small Search-LWE: Given A,c, find s,x • [R’05, P’09] As hard as worst-case of some lattice problems Learning with Errors (LWE) • n – security parameter • q poly(n) • m > n log q n s c =m c close to the linear code spanned by A A random mod q + x mod q small Decision-LWE: Distinguish c from random • [R’05] as hard as finding s,x • For certain parameters Learning with Errors (LWE) n m S C =m A n + random mod q • Many LWE instances with same A • Same hardness (easy hybrid argument) X small Ajtai’s Trapdoors • [A’96] Given A , hard to find small s.t. tA =0 mod q t • As hard as worst-case of some lattice problems • [A’99] But it is possible to generate together T A = 0 mod q small, full rank random • [Alwen-Peikert’08] Even smaller T Trapdoor Functions [GPV’08] • (A,s,x) As+x is a trapdoor function • Can use T to correct errors: • c = As + x • Tc = T(As + x) = Tx mod q • But T,x are small, so Tx << q (Tc mod q) = Tx • Equality over the integers T-1(Tc mod q) = x Our Cryptosystem Step 1: Encryption from linear ECCs • Code is the column space of A mod q • { As: s Zqn } • Bad representation (PK) is A itself • Given A, hard to distinguish words close to the code from random words (LWE) • Good representation (SK) is T • Can use T to correct errors Step 1: Encryption from linear ECCs • PK: A , SK: T • Encode plaintext is LSB of error matrix • Plaintext is a binary matrix Bmxm X • Enc(A,B): Choose random Smxn, small Emxm C = A S + X mod q 2E+B • Dec(T,C): Set X T-1(TC mod q) • Output B = X mod 2 Step 1: Encryption from linear ECCs • Security follows from LWE (for odd q) Thm: LWE For any B, EncA(B) random Proof: Given LWE input (A,C’) • Either C’=AS+E or C’ random: • Set C = 2C’+B mod q • If C’=AS+E then C = A(2S) + (2E+B) mod q • A random encryption of B • If C’ is random then so is C Step 1: Encryption from linear ECCs Additive homomorphism “for free” • C = C1 + C2 = (AS1+(2E1+B1)) + (AS2+(2E2+B2)) = A(S1+S2) + 2(E1+E2)+(B1+B2) mod q S X • T-1(TC mod q) = X = B1+B2 mod 2 • As long as X <<q Step 2: ECC lives inside a ring • Multiply C1 x C2 mod q? • (AS1+(2E1+B1)) (AS2+(2E2+B2)) = A(…) + (2E1+B1)AS2 + 2(…)+B1B2 mod q • Not what we wanted • Cannot use T to cancel out (2E1+B1)AS2 • Matrix multiplication is not commutative Step 2: ECC lives inside a ring • How about C = C1 x C2t mod q? • (AS1+(2E1+B1)) (AS2+(2E2+B2))t = A(…) + (…)At + 2(…)+B1B2t mod q X • That’s better: • TCTt = TXTt mod q • X = (2E1+B1)(2E2+B2)t is still small TCTt mod q = TXTt over the integers T-1(TCTt mod q)(Tt)-1 = X = B1B2t mod 2 What Did We Get? • • • • • KeyGen: Generate T A Enc(A, B): CAS + 2E+B mod q Add(C1,C2): CC1+C2 mod q Mult(C1,C2): CC1C2t mod q Dec(T, C): BT-1(TCTt mod q)(Tt)-1 mod 2 • Can decrypt any quadratic formula with polynomially many terms • With appropriate parameters What Did We Get? • • • • • Can KeyGen: Generate T A replace 2 Enc(A, B): CAS + pE+B mod q by any pq Add(C1,C2): CC1+C2 mod q Mult(C1,C2): CC1C2t mod q Dec(T, C): BT-1(TCTt mod q)(Tt)-1 mod p • Can decrypt any quadratic formula with polynomially many terms • With appropriate parameters Extensions, Applications • Can apply the [AMGH’10] transformation • Get homomorphism for low-degree polynomials • “Dual Regev encryption” [GPV’08] is a special case of our scheme* • Leakage resilience • IBE • Efficient quadratic-formula homomorphism for polynomials, big-integers * After changing encoding of plaintext Thank You 2-of-2 Decryption • • • • Alice has key-pair (A1,T1), Bob has (A2,T2) Charlie encrypts B1 to Alice, [ C1A1S1+X1 ]q Dora encrypts B2 to Bob, [ C2A2S2+X2 ]q Zachariah Sets C* = [ C1 C2t ]q • C* looks random to either Alice, Bob • Pulling their keys together they can recover B1B2t • B1B2t = T1-1[T1C*T2t]q (T2t)-1 mod 2 • Can also “blind” C* to hide relation to C1, C2 Multiplying Polynomials • p(x) = p0+p1x+p2x2, q(x) = q0+q1x+q2x2 p2 p1 p0 p2 p1 P= p2 q0 q1 q2 q0 q1 Q= q0 R= p0q1+p1q0+p1q0 p0q1+p1q0 p0q0 PQt+R= p1q2+p2q1 p2q2 $ $ $ $ $ $ $ $ Dual Regev Encryption [GPV’08] • Dual-Regev Cryptosystem is an instance -uof our scheme with T = 0 • A different input encoding than [GPV’08] • T is no longer invertible • But can still recover top-left entry in B • It is known to be IBE, leakage-resilient • Still true with new input encoding • And now it supports quadratic formulas