challenging research topic. At CRYPTO’19, Gohr proposed a Neural
Distinguisher (ND) based on a plaintext difference.
The ND takes a ciphertext pair as input and outputs its class (a real or random ciphertext pair).
At EUROCRYPTO’20, Benamira et al proposed a deeper analysis
of how two specific NDs against Speck32/64 work. However, there are
still three research gaps that researchers are eager to fill in.
(1) what features related to a ciphertext pair are learned by the ND?
(2) how to explain various phenomena related to NDs?
(3) what else can machine learning do in conventional cryptanalysis?
In this paper, we filled in the three research gaps: (1) we first propose
the Extended Differential-Linear Connectivity Table (EDLCT) which is
a generic tool describing a cipher. Features corresponding to the EDLCT
are designed to describe a ciphertext pair.
Based on these features, various machine learning-based distinguishers including the ND are built.
To explore various NDs from the EDLCT view, we propose a Feature Set
Sensitivity Test (FSST) to identify which features may have a significant influence on NDs.
Features identified by FSST share the same characteristic related to the cipher’s round function.
Surrogate models of NDs are also built based on identified features.
Experiments on Speck32/64 and DES confirm that features corresponding to the EDLCT are learned
by NDs. (2) We explain phenomena related to NDs via EDLCT.
(3) We show how to use machine learning to search differential-linear
propagations ∆ → λ with a high correlation, which is a tough task in the
differential-linear attack. Applications in Chaskey and DES demonstrate
the advantages of machine learning.
Furthermore, we provide some optional inputs to improve ND