Evaluating Deep Learning Techniques for Known-Plaintext Attacks on the Complete Columnar Transposition Cipher
Nino Fürthauer, Vasily Mikhalev, Nils Kopal, Bernhard Esslinger, Harald Lampesberger, Eckehard Hermann · Linköping electronic conference proceedings · 2022
This paper examines whether deep neural networks (DNN) can learn knownplaintext attacks on plaintext-ciphertextpairs, that were created by encrypting with complete columnar transposition. We propose a new algorithm that extends pure DNN-based prediction with additional post-processing steps to further enhance key prediction quality. Our approach is easily extensible and currently supports key lengths from 2 to 20 characters. Each key length has been empirically evaluated with plain-/ciphertextpairs of different lengths. For plain- and ciphertexts with a length of five times the key length, our algorithm achieves a success rate of 96% which is, to the best of our knowledge, a new state of the art on deep-learning-based known-plaintext attacks against columnar transposition.