Can Neural Networks spot the Sponge?

Abstract

Cryptographic hash functions ensure data integrity and security in modern cryptographic systems. They are a core mechanism for digital signatures, authentication protocols, and blockchain technology, designed to make their output impossible to predict. Inspired by machine learning’s ability to approximate complex functions, we trained models of varying complexity to assess whether they can learn and replicate the behavior of cryptographic hash functions. Our study evaluates the extent to which these models are able to approximate the underlying mathematical operations, shedding light on potential vulnerabilities and theoretical limits of machine learning in cryptographic contexts.