International Association
for Cryptologic Research

In this paper we present our solution to the CHES Challenge 2020, the task of which it was to break masked hardware respective software implementations of the lightweight cipher Clyde by means of side-channel analysis. We target the secret cipher state after processing of the first S-box layer. Using the provided trace data we obtain a strongly biased posterior distribution for the secret-shared cipher state at the targeted point; this enables us to see exploitable biases even before the secret sharing based masking. These biases on the unshared state can be evaluated one S-box at a time and combined across traces, which enables us to recover likely key hypotheses S-box by S-box.In order to see the shared cipher state, we employ a deep neural network similar to the one used by Gohr, Jacob and Schindler to solve the CHES 2018 AES challenge. We modify their architecture to predict the exact bit sequence of the secret-shared cipher state. We find that convergence of training on this task is unsatisfying with the standard encoding of the shared cipher state and therefore introduce a different encoding of the prediction target, which we call the scattershot encoding. In order to further investigate how exactly the scattershot encoding helps to solve the task at hand, we construct a simple synthetic task where convergence problems very similar to those we observed in our side-channel task appear with the naive target data encoding but disappear with the scattershot encoding.We complete our analysis by showing results that we obtained with a “classical” method (as opposed to an AI-based method), namely the stochastic approach, thatwe generalize for this purpose first to the setting of shared keys. We show that the neural network draws on a much broader set of features, which may partially explain why the neural-network based approach massively outperforms the stochastic approach. On the other hand, the stochastic approach provides insights into properties of the implementation, in particular the observation that the S-boxes behave very different regarding the easiness respective hardness of their prediction.

This paper has four main contributions. First, we calculate the predicted difference distribution of Speck32/64 with one specific input difference under the Markov assumption completely for up to eight rounds and verify that this yields a globally fairly good model of the difference distribution of Speck32/64. Secondly, we show that contrary to conventional wisdom, machine learning can produce very powerful cryptographic distinguishers: for instance, in a simple low-data, chosen plaintext attack on nine rounds of Speck, we present distinguishers based on deep residual neural networks that achieve a mean key rank roughly five times lower than an analogous classical distinguisher using the full difference distribution table. Thirdly, we develop a highly selective key search policy based on a variant of Bayesian optimization which, together with our neural distinguishers, can be used to reduce the remaining security of 11-round Speck32/64 to roughly 38 bits. This is a significant improvement over previous literature. Lastly, we show that our neural distinguishers successfully use features of the ciphertext pair distribution that are invisible to all purely differential distinguishers even given unlimited data.While our attack is based on a known input difference taken from the literature, we also show that neural networks can be used to rapidly (within a matter of minutes on our machine) find good input differences without using prior human cryptanalysis. Supplementary code and data for this paper is available at https://github.com/agohr/deep_speck.