Cryptology ePrint Archive: Report 2017/895

Linear Cryptanalysis of DES with Asymmetries

Andrey Bogdanov and Philip S. Vejre

Abstract: Linear cryptanalysis of DES, proposed by Matsui in 1993, has had a seminal impact on symmetric-key cryptography, having seen massive research efforts over the past two decades. It has spawned many variants, including multidimensional and zero-correlation linear cryptanalysis. These variants can claim best attacks on several ciphers, including PRESENT, Serpent, and CLEFIA. For DES, none of these variants have improved upon Matsui's original linear cryptanalysis, which has been the best known-plaintext key-recovery attack on the cipher ever since. In a revisit, Junod concluded that when using $2^{43}$ known plaintexts, this attack has a complexity of $2^{41}$ DES evaluations. His analysis relies on the standard assumptions of right-key equivalence and wrong-key randomisation.

In this paper, we first investigate the validity of these fundamental assumptions when applied to DES. For the right key, we observe that strong linear approximations of DES have more than just one dominant trail and, thus, that the right keys are in fact inequivalent with respect to linear correlation. We therefore develop a new right-key model using Gaussian mixtures for approximations with several dominant trails. For the wrong key, we observe that the correlation of a strong approximation after the partial decryption with a wrong key still shows much non-randomness. To remedy this, we propose a novel wrong-key model that expresses the wrong-key linear correlation using a version of DES with more rounds. We extend the two models to the general case of multiple approximations, propose a likelihood-ratio classifier based on this generalisation, and show that it performs better than the classical Bayesian classifier.

On the practical side, we find that the distributions of right-key correlations for multiple linear approximations of DES exhibit exploitable asymmetries. In particular, not all sign combinations in the correlation values are possible. This results in our improved multiple linear attack on DES using 4 linear approximations at a time. The lowest computational complexity of $2^{38.86}$ DES evaluations is achieved when using $2^{42.78}$ known plaintexts. Alternatively, using $2^{41}$ plaintexts results in a computational complexity of $2^{49.75}$ DES evaluations. We perform practical experiments to confirm our model. To our knowledge, this is the best attack on DES.

Category / Keywords: Linear cryptanalysis, DES, mixture models, right-key equivalence, wrong-key randomisation, linear hull, multiple linear

Original Publication (in the same form): IACR-ASIACRYPT-2017

Date: received 7 Sep 2017, last revised 18 Sep 2017

Contact author: anbog at dtu dk, psve@dtu dk

Available format(s): PDF | BibTeX Citation

Version: 20170918:130615 (All versions of this report)

Short URL: ia.cr/2017/895

Discussion forum: Show discussion | Start new discussion


[ Cryptology ePrint archive ]