Paper 2025/903

Rock and a Hard Place: Attack Hardness in Neural Network-assisted Side Channel Analysis

Abstract

Side-channel analysis (SCA) has become a crucial area in ensuring the security of hardware implementations against potential vulnerabilities. With advancements in machine learning (ML) and artificial intelligence (AI), neural network(NN)-assisted techniques for SCA have demonstrated significant effectiveness. However, a fundamental question remains unanswered: are traces corresponding to different subkeys equally hard to attack? This paper addresses this issue by leveraging explainable AI techniques to analyze the hardness levels and, consequently, the root cause of hardness. To systematically investigate this problem, we reintroduce hardness metrics in SCA, which have been known to the ML community. Those metrics include query hardness (QH), log odds (LO), and matrix-based Rényi entropy (MRE). The challenge in this regard is that (virtually all) hardness metrics in ML cannot be adopted as they are. This is because ML and SCA metrics have conflicting goals, namely boosting accuracy and rank. Through careful experimentation, we identify the shortcomings of QH and LO in SCA and recommend MRE as a suitable hardness metric for SCA. We also study how hardness has been seen in SCA, where recent work has suggested the influence of class “labels” on the attack difficulty. Employing rigorous evaluation, our paper demonstrates that no statistically significant evidence can be found to support this claim. This leads us to the question of how much traces’ time samples affect the inherent hardness in distinguishing key candidates. Our novel explainable AI (XAI) approach not only answers this, but also makes a link between hardness and rank as the common performance metric in SCA. Our findings indicate that hardness values are influenced mainly by time samples rather than specific key labels. Furthermore, we examine whether hardness captures intrinsic properties of the traces, specifically, their lack of separability in feature space due to their inherent similarities. To this end, we introduce, for the first time in the context of SCA, the use of maximum mean discrepancy (MMD) as a principled metric. MMD effectively links trace hardness with attack difficulty by quantifying distributional differences induced by traces’ time samples. In addition to visualization techniques showing the root cause of hardness based on MMD, we employ XAI to explain the connection between MMD and attack hardness. Our proposed methodology enhances the understanding of attack difficulty in SCA and contributes to developing more robust evaluation metrics for profiling attacks.