Paper 2021/1479

Reducing the Cost of Machine Learning Differential Attacks Using Bit Selection and aPartial ML-Distinguisher

Amirhossein Ebrahimi, Francesco Regazzoni, and Paolo Palmieri

Abstract

In a differential cryptanalysis attack, the attacker tries to observe a block cipher's behavior under an input difference: if the system's resulting output differences show any non-random behavior, a differential distinguisher is obtained. While differential cryptanlysis has been known for several decades, Gohr was the first to propose in 2019 the use of machine learning (ML) to build a distinguisher. In this paper, we present the first Partial Differential (PD) ML-distinguisher, and demonstrate its effectiveness on lightweight cipher SPECK32/64. As a PD-ML-distinguisher is based on a selection of bits rather than all bits in a block, we also study if different selections of bits have different impact in the accuracy of the distinguisher, and we find that to be the case. More importantly, we also establish that certain bits have reliably higher effectiveness than others, through a series of independent experiments on different datasets, and we propose an algorithm for assigning an effectiveness score to each bit in the block. By selecting the highest scoring bits, we are able to train a partial ML-distinguisher over 8-bits that is almost as accurate as an equivalent ML-distinguisher over the entire 32 bits (68.8% against 72%), for six rounds of SPECK32/64. The reduced input size implies a significant reduction in the complexity of achieving a distinguisher, and also leads to a reduction in the number of bits of possible subkeys to be guessed in a potential subsequent key recovery attack. These results may therefore open the way to the application of (partial) ML-based distinguishers to ciphers whose block size has so far been considered too large.