Multidimensional linear cryptanalysis

Abstract

Linear cryptanalysis is an important tool for studying the security of symmetric ciphers. In 1993 Matsui proposed two algorithms, called Algorithm 1 and Algorithm 2, for recovering information about the secret key of a block cipher. The algorithms exploit a biased probabilistic relation between the input and output of the cipher. This relation is called the (one-dimensional) linear approximation of the cipher. Mathematically, the problem of key recovery is a binary hypothesis testing problem that can be solved with appropriate statistical tools. The same mathematical tools can be used for realising a distinguishing attack against a stream cipher. The distinguisher outputs whether the given sequence of keystream bits is derived from a cipher or a random source. Sometimes, it is even possible to recover a part of the initial state of the LFSR used in a key stream generator. Several authors considered using many one-dimensional linear approximations simultaneously in a key recovery attack and various solutions have been proposed. In this thesis a unified methodology for using multiple linear approximations in distinguishing and key recovery attacks is presented. This methodology, which we call multidimensional linear cryptanalysis, allows removing unnecessary and restrictive assumptions. We model the key recovery problems mathematically as hypothesis testing problems and show how to use standard statistical tools for solving them. We also show how the data complexity of linear cryptanalysis on stream ciphers and block ciphers can be reduced by using multiple approximations. We use well-known mathematical theory for comparing different statistical methods for solving the key recovery problems. We also test the theory in practice with reduced round Serpent. Based on our results, we give recommendations on how multidimensional linear cryptanalysis should be used.