Paper 2020/436

Deep Learning based Side-Channel Attack: a New Profiling Methodology based on Multi-Label Classification

Houssem Maghrebi

Abstract

Deep Learning based Side-Channel Attacks (DL-SCA) are an emerging security assessment method increasingly being adopted by the majority of certification schemes and certification bodies to assess the resistance of cryptographic implementations. The related published investigations have demonstrated that DL-SCA are very efficient when targeting cryptographic designs protected with the common side-channel countermeasures. Furthermore, these attacks allow to streamline the evaluation process as the pre-processing of the traces (\emph{e.g.} alignment, dimensionality reduction, \dots) is no longer mandatory. In practice, the DL-SCA are applied following the divide-and-conquer strategy such that the target, for the training and the attack phases, only depends on $8$ key bits at most (to avoid high time complexity especially during the training). Then, the same process is repeated to recover the remaining bits of the key. To mitigate this practical issue, we propose in this work a new profiling methodology for DL-SCA based on the so-called multi-label classification. We argue that our new profiling methodology allows applying DL-SCA to target a bigger chunk of the key (typically $16$ bits) without introducing a learning time overhead and while guaranteeing a similar attack efficiency compared to the commonly used training strategy. As a side benefit, we demonstrate that our leaning strategy can be applied as well to train several intermediate operations at once. Interestingly, we show that, in this context, our methodology is even faster than the classical training and leads to a more efficient key recovery phase. We validated the soundness of our proposal on simulated traces and experimental data-sets; amongst them, some are publicly available side-channel databases. The obtained results have proven that our profiling methodology is of great practical interest especially in the context of performing penetration tests with high attack potential (\emph{e.g.} Common Criteria, EMVCO) where the time required to perform the attack has an impact on its final rating.