Paper 2026/1204

Robust Single-Trace Full-Key Extraction from Million-Point Traces With Cross-Implementation Transfer

Abstract

End-to-end deep-learning side-channel attacks on public-key implementations have recently become possible even for million-sample traces. However, existing methods require large computational resources and extract only partial key shares, which means that dedicated post-processing is required to turn detected leakage into demonstrations of successful key recovery attacks. We present an end-to-end sequence-to-sequence prediction approach to recover complete 256-bit key shares from single raw traces on the SCAAML ECC datasets recently studied by Bursztein et al (TCHES 2024). Our solution combines aggressive trace compression for dimensionality reduction with a 1-D U-Net trained using Connectionist Temporal Classification loss. The key idea is to decouple detecting leakage from mapping each leakage site to the correct part of the secret: the network outputs an annotated map of the trace marking likely leakage sites, and a greedy decoder reconstructs the ordered key bits from that map. Using synthetic tasks, we show that this division of labor circumvents a fundamental problem that causes neural network architectures and training methods commonly used in side-channel analysis to struggle with massive multi-target or misaligned extraction tasks. As a result, we are able to train a single extractor that achieves high accuracy on all four SCAAML ECC datasets in a single training run that takes minutes on a single GPU. The resulting extractors are robust, essentially maintaining their performance under large misalignment (we empirically tested rotations up to \(61\%\) of trace length), while degrading gracefully under a variety of trace corruptions, and even time reversal. They transfer across key shares and datasets with little degradation and no retraining. The U-Net outputs also yield prediction maps that localize leakage along the trace prior to decoding.