Differential Cryptanalysis for Multivariate Schemes

Transcription

1 Differential Cryptanalysis for Multivariate Schemes Jacques Stern Joint work with P. A. Fouque and L. Granboulan École normale supérieure Differential Cryptanalysis for Multivariate Schemes p.1/23

2 MI Cryptosystem F q a finite field of characteristic 2 Secret Key : S, T two affine bijections in (F q ) n F is defined as F (X) = X ql +1 in F q n and is thus a quadratic map from (F q ) n to (F q ) n Public key : the system E of equations in (F q ) n E = T F S Decryption function : invert T, compute F 1 by raising to the power (q l + 1) 1 mod (q n 1), and invert S Differential Cryptanalysis for Multivariate Schemes p.2/23

3 Perturbated MI Cryptosystem (PMI) R linear map from (F q ) n to (F q ) r with r n H quadratic function from (F q ) r to (F q ) n E = T (F + H R) S = E + T H R S The PMI scheme E is the MI scheme E plus a random-looking quadratic term T H R S q r must be small so that exhaustive search on q r is efficient, otherwise decryption is slow Secret key : (S, T, P ) where P is a table storing (λ, µ) pairs s.t. H(µ) = λ Differential Cryptanalysis for Multivariate Schemes p.3/23

4 MI and PMI Cryptosystems Message x Message x A S 1 F 1 E F u + A S µ λ R H B T 1 Ciphertext y B T Ciphertext y Differential Cryptanalysis for Multivariate Schemes p.4/23

5 PMI Decryption Algorithm Input : y ciphertext Output : x plaintext s.t. y = E (x) Compute B = T 1 (y) For the q r pairs (λ, µ), compute A λ = F 1 (B λ) until R(A λ ) = µ Return x λ = S 1 (A λ ) If many pairs (λ, µ) are possible, redundancy is added to the plaintext Differential Cryptanalysis for Multivariate Schemes p.5/23

6 PMI schemes and variants Ding s practical cryptosystem q = 2, n = 136, l = 40 and r = 6 so F (X) = X , R : (F 2 ) 136 (F 2 ) 6 and H : (F 2 ) 6 (F 2 ) 136 gcd( , ) = 2 gcd(136,40) 1 = The variant of PMI when gcd(n, l) = 8 is called Ding s scheme The variant of PMI when gcd(n, l) = 1 is called Generalized scheme Differential Cryptanalysis for Multivariate Schemes p.6/23

7 Patarin attack on MI Search n bilinear relations (B i ) 1 i n between the plaintext x and the ciphertext y Recover the coefficients of the bilinear relations using O(n 2 ) plaintext/ciphertext pairs Given a ciphertext y, solve the system of the n bilinear relations to find the plaintext x However, the system is not invertible ( exhaustive search to uniquely recover x) Differential Cryptanalysis for Multivariate Schemes p.7/23

8 Patarin attack on (2) Let A = S(x) F q n and B = T 1 (y) F q n Since F (A) = B, we have B = A ql +1 By raising to the power q l 1 and multiplying by AB, we get a bilinear expression A B ql = A q2l B Rewriting this equation in the variables x and y and projeting into (F q ) n, we get n bilinear relations between the plaintext and ciphertext Differential Cryptanalysis for Multivariate Schemes p.8/23

9 Breaking the PMI scheme E = E + T H R S Here, constants of affine maps are erased (see paper) If k K = ker(r S), then E (k) = E(k) On the subspace K, Patarin s attack can be applied Goal : decrypting all PMI ciphertexts when x K whose dimension (n r) is large for all x Detecting membership in K using differential cryptanalysis Differential Cryptanalysis for Multivariate Schemes p.9/23

10 The use of differentials Let G be a quadratic map, its differential is linear L G,k : x G(x + k) G(x) G(k) + G(0) The constant term disappears thanks to G(0), and so L G,k is a linear map and not an affine one Let X = S(x) and K = S(k) Differential of a composition of functions : if E = T F S, then L E,k (x) = T L F,K (X) Since S and T are bijection, dim(ker(l E,k )) = dim(ker(l F,K )) Differential Cryptanalysis for Multivariate Schemes p.10/23

11 Expression of L F,K L F,K (X) = F (X + K) F (X) F (K) + F (0) = (X + K) ql (X + K) X ql +1 K ql +1 = (X ql + K ql ) (X + K) X ql +1 K ql +1 ( ( ) ) q l = K ql X + X ql K = K ql +1 X X K + K X L F,K (X) is a linear map Differential Cryptanalysis for Multivariate Schemes p.11/23

12 Kernel s dimension of the differential in MI X is in the kernel of L F,K L F,K (X) = 0 Y + Y ql = 0 where Y = X K Y (1 + Y ql 1 ) = 0 Y ql 1 = 1 since char(f q ) = 2 Y = 1 K ker L F,K k ker L E,k The equation Y ql 1 = 1 has q gcd(l,n) 1 solutions Therefore, dim(ker L E,k ) = dim(ker L F,K ) = gcd(l, n) Differential Cryptanalysis for Multivariate Schemes p.12/23

13 Kernel s dimension of the differential in PMI What is the contribution of H R on the kernel s dimension? Since H is quadratic, its differential is L H R,K (X) = r i,j=1 α i,j[r i (X)R j (K) + R i (K)R j (X)] K is always in ker(l H R,K ) and dim(ker(l E,K)) 1 Since H is random, L H R,K is a random linear map and L E,k is also a random linear map Consequently, dim(ker L E,k) follows the distribution of random linear map Differential Cryptanalysis for Multivariate Schemes p.13/23

14 Breaking Ding s scheme In the proposed system, gcd(l, n) = 8 The probability that a linear map has a kernel of dimension 8 is small ( 1/2 20 ) We devise the following test : if dim(ker(l E,k)) = gcd(l, n), then decide k K otherwise decide k K Differential Cryptanalysis for Multivariate Schemes p.14/23

15 Total Break of Ding s scheme K can be recovered by collecting n r independent vectors as well as the bilinear relations of Patarin s attack when k K On this subspace, we can invert any ciphertext y s.t. x K where y = E (x) which holds with probability 1/q r The entire space can be divided into q r affine subspaces parallel to the K direction The same attack can be mounted in parallel on all these subspaces to recover any ciphertext y Differential Cryptanalysis for Multivariate Schemes p.15/23

16 Breaking the Generalized scheme When gcd(l, n) = 1, the previous test cannot be applied since dim(ker L E,k) = gcd(l, n) = 1 with high probability even if k K Therefore, if dim(ker L E,k) = 1, k may or not be in K if dim(ker L E,k) > 1, k K with probability 1 We need to filter bad values k s.t. dim(ker L E,k) = 1 and k K Differential Cryptanalysis for Multivariate Schemes p.16/23

17 Filtering the bad values k Since K is a linear space, if k, k K, then k + k K To decide if k K, which holds with probability 1/q r, take different k s.t. dim(ker L E,k ) = 1 and compute the distribution of dim(l E,k+k ) The distributions of dim(l E,k+k ) when k K and when k K are different and can be distinguished by statistic experiments Differential Cryptanalysis for Multivariate Schemes p.17/23

18 New Attack on the MI cryptosystem This new attack finds two bilinear relations C and D of n coordinates : C is between a vector f k of the kernel of the transpose matrix of L E,k and the ciphertext y corresponding to E(k) D is between the vector f k and the corresponding plaintext k Differential Cryptanalysis for Multivariate Schemes p.18/23

19 Decomposition of L E,k ( Since L F,K (X) = K ql +1 X + ( ) ) X q l, K K L E,k = T L F,K S can be written as T µ K ψ θ K S where µ K, ψ and θ K are the linear maps and K = S(k) and X = S(x) : θ K : X X K ψ : Y Y + Y ql independent of K µ K : Z K ql +1 Z Differential Cryptanalysis for Multivariate Schemes p.19/23

20 f k in the kernel of transpose of L E,k T, µ K, ψ, θ K and S are n n matrices, and (f k ) is a row vector in L E,k s.t. (f k )(T.µ K.ψ.θ K.S) = 0 Since θ K and S invertible matrices, (f k )(T.µ K ) ker ψ If gcd(l, n) = 1, then dim(ker ψ) = 1 and if q = 2 (f k )(T.µ K ) = ( ˆf) Differential Cryptanalysis for Multivariate Schemes p.20/23

21 The two bilinear relations C and D µ K (Z) = F (K) Z is linear in F (K) Since F (K) = T 1 (E(k)), then µ K is linear in the ciphertext E(k) So (f k )(T.µ K ) = ( ˆf) is a bilinear relation C between E(k) and f k which can be projected to the n coordinates Finally, as (f k )(L E,k ) = 0 and L E,k is linear in k, then there is a bilinear relation D between f k and the plaintext k Differential Cryptanalysis for Multivariate Schemes p.21/23

22 The new attack against MI Precomputation stage : Using many plaintexts k, compute f k (kernel of L E,k ) and the corresponding ciphertexts E(k) and recover the bilinear relations C(f k, E(k)) recover the bilinear relations D(f k, k) On-line stage : Given a ciphertext E(k), recover the vector f k using C and decrypt using D and f k Differential Cryptanalysis for Multivariate Schemes p.22/23

23 Conclusion We show that differential cryptanalysis is a nice tool which can be adapted to successfully attack multivariate schemes We apply this novel cryptanalytic method in order to propose A new attack against the MI original scheme An attack against a recently proposed variant of MI called PMI Differential Cryptanalysis for Multivariate Schemes p.23/23