Paper 2024/1437

HierNet: A Hierarchical Deep Learning Model for SCA on Long Traces

Abstract

In Side-Channel Analysis (SCA), statistical or machine learning methods are employed to extract secret information from power or electromagnetic (EM) traces. In many practical scenarios, raw power/EM traces can span hundreds of thousands of features, with relevant leakages occurring over only a few small segments. Consequently, existing SCAs often select a small number of features before launching the attack, making their success highly dependent on the feasibility of feature selection. However, feature selection may not always be possible, such as in the presence of countermeasures like masking or jitters. Several recent works have employed deep learning (DL) methods to conduct SCA directly on long raw traces, thereby reducing reliance on feature selection steps. However, these methods often perform poorly in the presence of various jitter-based countermeasures. While some methods, such as the one proposed by Hajra et al. (TCHES 2024), have shown high robustness to several jitter-based countermeasures on relatively short traces, we demonstrate in this work that their performance deteriorates as trace lengths increase. Consequently, existing DL models perform poorly on long traces or in the presence of various jitter-based countermeasures. To address these challenges, we propose HierNet, a hierarchical DL model designed for SCA on long traces. HierNet uses a two-level information assimilation process: at the base level, a shift-invariant DL model processes smaller trace segments; at the top level, another DL model aggregates the outputs from all segments to produce the final output. HierNet has been experimentally evaluated against various combinations of masking, random delay, and clock jitter countermeasures, using three publicly available SCA datasets with trace lengths up to 250K features. The results have been compared with four existing SCA benchmark models. They demonstrate HierNet's superiority, particularly on long traces or in the presence of clock jitter countermeasures, showcasing its ability to reach guessing entropy 1 with fewer than or around 10 attack traces, while the benchmark models fail to do so even using 5K attack traces. HierNet also exhibits significantly better results in low training data scenarios.