Machine Learning Attacks On SPECK

Despite only a recent start, machine learning (ML) has gained traction in cryptographic analysis, more specifically in the realm of differential distinguishers for symmetric key ciphers. Based on the previous ML based models on lightweight ciphers presented by Baksi et al. (Eprint’20/DATE’21) and Baksi et al. (SILC’21), we choose the lightweight block cipher family SPECK as the target for our analysis in this paper. We show new results on SPECK-32 and SPECK-128 using two well-known ML libraries, TensorFlow (with Keras API) and PyTorch. We experiment with various options for the architecture (such as number of layers, dropout and activation function). Among other results, we present a differential distinguisher for 6 rounds of SPECK-32 and a strong indication for it to follow through one more round, with trivial complexity, which complements the analysis done by Gohr (CRYPTO’19).