Linear Attack on Round-Reduced DES Using Deep Learning

Linear attack is a powerful known-plaintext cryptanalysis method on block ciphers, which has been successfully applied in DES, KATAN, SPECK and other ciphers. In this paper, we use deep learning networks to achieve linear attack on DES with plain-cipher pairs. Comparing with traditional linear attack algorithm, our work requires less knowledge about complex cryptanalysis as neural network can work well by data-driven. Thus, this paper has three main contributions. First, a new linear attack architecture based on deep residual network was proposed to train discriminative neural networks with auto-generated plain-cipher pair data. The results indicate that trained neural networks can effectively learn algorithmic representations of the XOR distributions of given linear expression on DES. Second, several novel neural network-based algorithms were designed to efficiently enforce key recovery on round-reduced DES using trained networks with moderate full and partial bits of linear expression as inputs. Third, as far as we know, it is the first time that neural networks are used to achieve known-plaintext attack on complex block ciphers.