Paper 2017/204

Linear Cryptanalysis Using Low-bias Linear Approximations

Tomer Ashur, Daniël Bodden, and Orr Dunkelman

Abstract

This paper deals with linear approximations having absolute bias smaller than $2^{-\frac{n}{2}}$ which were previously believed to be unusable for a linear attack. We show how a series of observations which are individually not statistically significant can be used to create a $\chi^2$ distinguisher. This is different from previous works which combined a series of significant observations to reduce the data complexity of a linear attack. We test the distinguisher on a real-world cipher and show that it can be used to improve previous results.