Cryptanalysis of Round-Reduced ASCON powered by ML

Abstract

Our research focuses on attacking Ascon, a lightweight block cipher presented as a candidate in the NIST Lightweight Cryptography Standardization Process. This block cipher provides authenticated encryption with associated data functionalities. We propose a cryptanalysis model based on deep learning (DL), where the goal is to predict plaintext bits given knowledge of the ciphertext and other publicly known cipher input parameters. Our experiments show that such knownplaintext attacks can be successfully executed on a round reduced version of the cipher stripped of the finalization phase. This, in turn, validates the theoretical results. Cryptographic algorithms are complex for the purpose of security and cannot be easily broken by an ML model in their regular form (not reduced). We explore multiple dataset generation techniques, model design, and training hyperparameters.