On the provable security of block ciphers and its impact on data-driven cryptanalysis

Established paradigms for block cipher designs encompass the generalized Feistel network, substitution-permutation network, and add-rotate-XOR (ARX) constructions. In the past, researchers have investigated the provable security guarantees provided by these classical design paradigms. Recently, significant attention has turned toward machine or deep learning-based cryptanalysis of block ciphers, aiming to train models capable of distinguishing between a cipher's output and that of a random function. Several of these approaches have outperformed classical cryptanalysis methods. This project investigates the relationship between a block cipher's provable security, particularly in terms of its indifferentiability from an ideal cipher, and its resilience against data-driven distinguishing attacks. The findings will contribute to a deeper understanding of these attacks, which are predominantly perceived as black box methodologies.