On possibility of using convolutional neural networks for creating universal attacks on iterative block ciphers

The paper explores possibility of applying convolutional neural networks to the security analysis of iterative block ciphers. A new approach for constructing distinguishing attacks based on a convolutional neural network is proposed. The approach is based on distinguishing between graphic equivalents of ciphertexts received by the CTR (counter) encryption mode after different number of rounds, including the number of rounds guaranteeing satisfaction of statistical properties. Several schemes are presented for constructing distinguishing attacks, which in some cases make it possible to detect deviations from randomness in smaller samples than previously known, and with a large number of rounds. The approach allows to create distinguishers without the need for an analytical research of each cipher, which makes it possible to build universal distinguishers for a series of ciphers.