Paper 2025/1251

Black Box to Blueprint: Visualizing Leakage Propagation in Deep Learning Models for SCA

Abstract

Deep learning (DL)-based side-channel analysis (SCA) has emerged as a powerful approach for extracting secret information from cryptographic devices. However, its performance often deteriorates when targeting implementations protected by masking and desynchronization-based countermeasures, or when analyzing long side-channel traces. In earlier work, we proposed EstraNet, a Transformer Network (TN)-based model designed to address these challenges by capturing long-distance dependencies and incorporating shift-invariant attention mechanisms. In this work, we perform an in-depth analysis of the internal behavior of EstraNet and propose methods to further enhance its effectiveness. First, we introduce {\bf DL-ProVe} (Deep Learning Leakage Propagation Vector Visualization), a novel technique for visualizing how leakage from secret shares in a masked implementation propagates and recombines into the unmasked secret through the layers of a DL model trained for SCA. We then apply DL-ProVe to EstraNet, providing the first detailed explanation of how leakage is accumulated and combined within such an architecture. Our analysis reveals a critical limitation in EstraNet’s multi-head GaussiP attention mechanism when applied to long traces. Based on this insights, we identify a new architectural hyperparameter which enables fine-grained control over the initialization of the attention heads. Our experimental results demonstrate that tuning this hyperparameter significantly improves EstraNet’s performance on long traces (with upto 250K features), allowing it to reach the guessing entropy 1 using only 3 attack traces while the original EstraNet model fails to do so even with 5K traces. These findings not only deepen our understanding of EstraNet’s internal workings but also introduce a robust methodology for interpreting, diagnosing, and improving DL models for SCA.