A Massive Machine-Learning Approach For Classical Cipher Type Detection Using Feature Engineering

Abstract

Cryptanalysis of enciphered documents typically starts with identifying the cipher type. A large number of encrypted historical documents exists, whose decryption can potentially increase the knowledge of historical events. This paper investigates whether machine learning can support the cipher type classification task when only ciphertexts are given. A selection of engineered features for historical ciphertexts and various machine-learning classifiers have been applied for 56 different cipher types specified by the American Cryptogram Association. Different neuronal network models were empirically evaluated. Our best-performing model achieved an accuracy of 80.24% which improves the current state of the art by 37%. Accuracy is calculated by dividing the total number of samples by the number of true positive predictions. The software-suite is published under the name "Neural Cipher Identifier (NCID)".

Proceedings of the 4th International Conference on Historical Cryptology HistoCrypt 2020

“…Finally The present study is not the first to apply classification or clustering methods to classical cryptography. For example, the Neural Cipher Identifier (NCID) by Leierzopf et al (2021) uses an ensemble neural network classifier for 55 standardized classical cipher types. More relevantly, Juzek (2019) clustered true ciphers and pseudociphers using support-vector machines on entropy.…”

The cipher of Emperor Rudolf II’s “Alchemical Hand Bell”

“…These basic observations are a good starting point to try to diagnose the cipher type (Callimahos, 1977, Chapter XI). Two diagnosis tools, based on ACA ciphers and using machine learning techniques are currently available online: Mason (BION) 8 and the tool from Leierzopf et al (2021) known as "NCID" 9 . Both tools provide a "probability" score in percentage terms ranking various possible ciphers.…”