International Association
for Cryptologic Research

Type-II Generalized Feistel Schemes are one of the most popular versions of Generalized Feistel Schemes. Their round function consists in applying a classical Feistel transformation to p sub-blocks of two consecutive words and then shifting the k = 2p words cyclically. The low implementation costs it offers are balanced by a low diffusion, limiting its efficiency. Diffusion of such structures may however be improved by replacing the cyclic shift with a different permutation without any additional implementation cost. In this paper, we study ways to determine permutations with the fastest diffusion called optimal permutations.To do so, two ideas are used. First, we study the natural equivalence classes of permutations that preserve cryptographic properties; second, we use the representation of permutations as coloured trees.For both heuristic and historical reasons, we focus first on even-odd permutations, that is, those permutations for which images of even numbers are odd. We derive from their structure an upper bound on the number of their equivalence classes together with a strategy to perform exhaustive searches on classes. We performed those exhaustive searches for sizes k ≤ 24, while previous exhaustive searches on all permutations were limited to k ≤ 16. For sizes beyond the reach of this method, we use tree representations to find permutations with good intermediate diffusion properties. This heuristic leads to an optimal even-odd permutation for k = 26 and best-known results for sizes k = 64 and k = 128.Finally, we transpose these methods to all permutations. Using a new strategy to exhaust equivalence classes, we perform exhaustive searches on classes for sizes k ≤ 20 whose results confirmed the initial heuristic: there always exist optimal permutations that are even-odd and furthermore for k = 18 all optimal permutations are even-odd permutations.

We consider highly structured truncated differential paths to mount a new rebound attack on Grøstl-512, a hash functions based on two AES-like permutations, P1024 and Q1024, with non-square input and output registers. We explain how such differential paths can be computed using a Mixed-Integer Linear Programming approach. Together with a SuperSBox description, this allows us to build a rebound attack with a 6-round inbound phase whereas classical rebound attacks have 4-round inbound phases. This yields the first distinguishing attack on a 11-round version of P1024 and Q1024 with about 272 computations and a memory complexity of about 256 bytes, to be compared with the 296 computations required by the corresponding generic attack. Previous best results on this permutation reached 10 rounds with a computational complexity of about 2392 operations, to be compared with the 2448 computations required by the corresponding generic attack.

A good linear diffusion layer is a prerequisite in the design of block ciphers. Usually it is obtained by combining matrices with optimal diffusion property over the Sbox alphabet. These matrices are constructed either directly using some algebraic properties or by enumerating a search space, testing the optimal diffusion property for every element. For implementation purposes, two types of structures are considered: Structures where all the rows derive from the first row and recursive structures built from powers of companion matrices. In this paper, we propose a direct construction for new recursive-like MDS matrices. We show they are quasi-involutory in the sense that the matrix-vector product with the matrix or with its inverse can be implemented by clocking a same LFSR-like architecture. As a direct construction, performances do not outperform the best constructions found with exhaustive search. However, as a new type of construction, it offers alternatives for MDS matrices design.