Abstract
Hashing is a crucial step in the Digital Signature (DS) generation and verification process. Traditional DSs rely on mathematically-defined hashing algorithms and public-private key encryption that offer strong formal guarantees. Our previous work introduces ML-256, a deep hashing model trained on random input-output pairings to produce a fixed-length hash-like output. That work demonstrates the feasibility of integrating Machine Learning (ML) based hashing into DS generation and highlights potential benefits in adaptability and maintainability. In this paper, we evaluate ML-256 through a more rigorous empirical evaluation of its behavior as a hash-like mapping algorithm. Rather than claiming formal cryptographic guarantees, we assess ML-256 through empirical tests of determinism, collision behavior under finite sampling, and resistance inversion using ML-based reconstruction attacks. This work additionally analyzes the computational trade-offs introduced by an ML approach by comparing throughput, DS generation and verification latency, and memory consumption against traditional DS pipelines. Results show that while ML-256 is slower and more resource expensive than traditional cryptographic functions, its outputs are difficult to invert using the tested reconstruction methods and exhibit no collisions within the evaluated regimes. Our findings demonstrate that robustness against ML-based attacks does not imply suitability as a cryptographic hash for DS. However, these results hint that ML-256 could be experimentally robust and easily configurable hash-line function for DS.