The past ten years have seen an increasing presence of deep learning algorithms to perform a diverse set of tasks thanks to advances in technology. A few years ago, those algorithms started to be used to help perform profiled side-channel analysis. Side-Channel Analysis (SCA) is a type of attack against secure algorithms which uses leakages of information contained in physical values, such as the power consumption, in order to retrieve the secret or a part of the secret used during the encryption. Profiled SCA corresponds to attacks where the attacker has access to a device similar to the one under attack and this allows him to build a profile, or template, of the physical quantity he is monitoring. Attacks using deep learning algorithms to build the profiles, commonly refer to as DLSCA, have been shown to perform as well, or even better, as the most powerful profiled attack, templates attacks. They also benefit from having higher resilience against classical countermeasures such as desynchrinization and masking.

On the other hand, this new method brings other challenges as the training of the networks determines their efficiency and thus is of high importance. In this presentation, after an introduction to DLSCA, we will explore how to evaluate the networks and their training in a side-channel context and how to improve the performances of basic networks.