Polyalphabetic cipher decryption function learning with LSTM networks

Oxford, United Kingdom

2024-06-26

While Recurrent Neural Networks have been applied to a wide range of problems, from language modelling to time series forecasting among many others, the possibility of approximating a decryption function from a machine producing pseudorandom sequences seems intuitively not something they would be good at. However, we show how LSTM networks are indeed capable of not only learning but also extracting external key information from known-plaintexts of only 15 characters in length. In order to do that, we model and simulate simpler ciphers such as the Vigenère and the Playfair along with more complex machines like the Siemens and Halske T52d (without KTF) and the Hagelin C-38 in which we train our networks with. Furthermore, we also analyse the effects of different input types such as randomized data, German literature and war intercepts decrypted by the FRA in Sweden. As a result, this approach has proven to be effective for the Vigenère and the C-38 as well as partially for the T52d while giving negative results for the Playfair. Although this is without doubt not better than preexisting techniques, the intention is not to describe it as a better method to extract the key of a given ciphertext but rather to demonstrate the potential of other non-standard approaches to accomplishing such tasks which can be said of being completely possible.