Paper 2026/338

Is it Really Broken? The Failure of DL-SCA Scoring Metrics under Non-Uniform Priors

Abstract

This paper investigates a recent, claimed state-of-the-art attack on the ASCADv2 dataset, a higher-order masked and shuffled AES implementation, which we demonstrate to be a false positive. Despite successful validation using classical metrics, including a converging Guessing Entropy (GE), we prove that the model learned no actual side-channel leakage. Instead, it exploited a statistical bias in the intermediate value distribution. We argue that the usual scoring function used in the GE is an unreliable metric in the presence of such biases. To address this critical evaluation flaw, we propose a set of methods to avoid falling in this pitfall. First, we introduce pre-emptive methods to detect significant biases in the target's value distribution before profiling, as well as post-mortem ones to examine the resulting model. Second, we present guidelines to avoid regimes where the GE is unreliable, and we derive the Asymptotically Optimal Distinguisher, a new, lightweight distinguisher that provably neutralizes the influence of learned priors in the GE metric, thereby isolating the information gained purely from the side-channel leakage. We demonstrate our methodology by successfully identifying the ASCADv2 false positive and applying it to synthetically biased versions of the ASCADv1 dataset.