Furthermore, with an ensemble model of feature-engineering and feature-learning neural network types, 82.78% accuracy over the same dataset has been achieved, which is the best known result for this significant problem in the field of cryptanalysis. Over a dataset of 10 million ciphertexts with a text length of 100 characters, Transformer correctly identified 72.33% of the ciphers, which is a slightly worse result than the best feature-engineering approach.

To lower the necessary data and the training time, various transfer-learning scenarios are investigated. To overcome this dependency on domain knowledge, we explore in this paper the applicability of the two feature-learning algorithms long short-term memory (LSTM) and Transformer, for 55 classical cipher types from ACA. Identifying these cipher types using machine learning has shown promising results, but the state of the art relies on engineered features based on cryptanalysis. The American Cryptogram Association (ACA) standardized so-called classical ciphers, which had historical relevance up to World War II. Cryptanalysis has acquired deep knowledge on cryptographic weaknesses of classical ciphers, and modern ciphers have been designed to circumvent these weaknesses. To break a ciphertext, as a first step, it is essential to identify the cipher used to produce the ciphertext.

Cipher identifier