Improved Neural Differential Distinguisher Model for Lightweight Cipher Speck

At CRYPTO 2019, Gohr proposed the neural differential distinguisher using residual network structure in convolutional networks on round-reduced Speck32/64. In this paper, we construct a 7-round for Speck32/64, which results better than Gohr’s work. The details are as follows. Firstly, new data format (C_r,C_r′,d_l,Cl,Cr,Cl′,Cr′) is input of distinguisher, can help to identify features previous round ciphertexts Speck algorithm. Secondly, paper modifies convolution layer block network, inspired by Inception module GoogLeNet. For experiments show that accuracy 97.13%, 9.1% and also higher currently known 89.63%. improve 2.38% 2.1%, respectively. Finally, demonstrate effectiveness key recovery attack performed 8-rounds success rate recovering correct 92%, with no more two incorrect bits. briefly discussed effect number ciphertext pairs sample training distinguisher. When total kept constant, increases s, but it leads occurrence overfitting.