Deep neural networks aiding cryptanalysis: A case study of the Speck distinguisher

At CRYPTO\u2719, A. Gohr proposed neural distinguishers for the lightweight block cipher Speck32/64, achieving better results than the state-of-the-art at that point. However, the motivation for using that particular architecture was not very clear, leading us to investigate whether a smaller and/or better performing neural distinguisher exists. This paper studies the depth-10 and depth-1 neural distinguishers proposed by Gohr with the aim of finding out whether smaller or better-performing distinguishers for Speck32/64 exist. We first evaluate whether we can find smaller neural networks that match the accuracy of the proposed distinguishers. We answer this question in affirmative with the depth-1 distinguisher successfully pruned, resulting in a network that remained within one percentage point of the unpruned network\u27s performance. Having found a smaller network that achieves the same performance, we examine if its performance can be improved as well. We also study whether processing the input before giving it to the pruned depth-1 network would improve its performance. To this end, convolutional autoencoders were found that managed to reconstruct the ciphertext pairs successfully, and their trained encoders were used as a preprocessor before training the pruned depth-1 network. We found that, even though the autoencoders achieve a perfect reconstruction, the pruned network did not have the necessary complexity anymore to extract useful information from the preprocessed input, motivating us to look at the feature importance to get more insights. To achieve this, we used LIME, with results showing that a stronger explainer is needed to assess it correctly