Can Sequence-to-Sequence Models Crack Substitution Ciphers?

Abstract

Decipherment of historical ciphers is a challenging problem. The language of the target plaintext might be unknown, and ciphertext can have a lot of noise. State-of-the-art decipherment methods use beam search and a neural language model to score candidate plaintext hypotheses for a given cipher, assuming the plaintext language is known. We propose an end-to-end multilingual model for solving simple substitution ciphers. We test our model on synthetic and real historical ciphers and show that our proposed method can decipher text without explicit language identification while still being robust to noise.

Decipherment as Regression: Solving Historical Substitution Ciphers by Learning Symbol Recurrence Relations

“…Seq2Seq Following (Aldarrab and May, 2021), this is a character level Transformer architecture that is only optimized on the target-side (plaintext) loss: Target-Only CausalLM is only optimized on the target-side loss L T GT , and incurs no loss when generating the source text.…”

“…For simple substitution ciphers, we use the same English data as above to create 1.2M synthetic substitution ciphers with lengths up to 256. Following previous work on 1:1 ciphers (Nuhn et al, 2013;Aldarrab and May, 2021), we evaluate on 50 test ciphers of lengths up to 128 (16,32,64) and beyond 128 (128,256) from the Wikipedia page on History 5 . All our experimental settings include data with word boundaries denoted by the space symbol (_).…”

“…Evaluation Following prior work Aldarrab and May, 2021), we evaluate on Symbol Error Rate (SER), the proportion of ciphertext symbols which are wrongly recovered.…”

“…CipherGAN (Gomez et al, 2018) exploits learned letter embedding distributions, but requires a large volume of ciphertext and only handles 1:1 substitution and Vigenère ciphers. Luo et al (2021) and Aldarrab and May (2022) pro-pose techniques to decipher undersegmented ciphers. Aldarrab and May (2021) train a sequenceto-sequence neural translation model to decipher from character frequencies.…”

“…Automated computational decipherment of such texts is challenging (Pettersson and Megyesi, 2019;Megyesi et al, 2020). Prior work has mainly focused on using clever heuristics and/or search algorithms to explore the space of cipher keys and score multiple candidate plaintexts under character language models (LMs) (Knight et al, 2006;Corlett and Penn, 2010;Hauer et al, 2014;Berg-Kirkpatrick and Klein, 2013;Nuhn et al, 2013Nuhn et al, , 2014 In contrast Aldarrab and May (2021) train a sequence-tosequence model to solve simple (one-to-one) substitution ciphers. This approach, however, cannot solve complex homophonic ciphers as it relies on frequency information which such ciphers obscure.…”