International Association
for Cryptologic Research

In this paper we provide new theoretical and empirical evidences that gradient-based deep learning profiling attacks (DL-SCA) suffer from masking schemes. This occurs through an initial stall of the learning process: the so-called plateau effect. To understand why, we derive an analytical expression of a DL-SCA model targeting simulated traces which enables us to study an analytical expression of the loss. By studying the loss landscape of this model, we show that not only do the magnitudes of the gradients decrease as the order of masking increases, but the loss landscape also exhibits a prominent saddle point interfering with the optimization process. From these observations, we (1) propose the usage of a second-order optimization algorithm mitigating the impact of low-gradient areas. In addition, we show how to leverage the intrinsic sparsity of valuable information in SCA traces to better pose the DL-SCA problem. To do so, we (2) propose to use the implicit regularization properties of the sparse mirror descent. These propositions are gathered in a new publicly available optimization algorithm, Scoop. Scoop combines second-order derivative of the loss function in the optimization process, with a sparse stochastic mirror descent. We experimentally show that Scoop pushes further the current limitations of DL-SCA against simulated traces, and outperforms the state-of-the-art on the ASCADv1 dataset in terms of number of traces required to retrieve the key, perceived information and plateau length. Scoop also performs the first non-worst-case attack on the ASCADv2 dataset. On simulated traces, we show that using Scoop reduces the DL-SCA time complexity by the equivalent of one masking order.