Paper 2024/1437

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

Abstract

Side-channel analysis (SCA) compromises the security of cryptographic devices by exploiting various side-channel leakages such as power consumption, electromagnetic (EM) emanations, or timing variations, posing a practical threat to the security and privacy of modern digital systems. In power or EM SCA, statistical or machine learning methods are employed to extract secret information from power/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 on long raw traces, thereby reducing dependence on feature selection steps. However, these methods often perform poorly against various jitter-based countermeasures. While some of these methods have shown high robustness to jitter-based countermeasures on relatively shorter traces, we demonstrate in this work that their performance deteriorates as trace lengths increase. Based on these observations, we develop a hierarchical DL model for SCA on long traces that is robust against various countermeasures. The proposed model, HierNet, extracts information from long traces using a two-level information assimilation process. At the base level, a DL model with shift-invariance is employed to extract information from smaller trace segments. Subsequently, a top-level DL model integrates the outputs of the base model to generate the final output. The proposed model has been experimentally evaluated against various combinations of masking, random delay, and clock jitter countermeasures using traces with lengths exceeding $200K$ features. The results have been compared with three existing SCA benchmark models. They demonstrate HierNet's superiority in several scenarios, such as on long traces, against clock jitter countermeasures, and low training data scenarios. In particular, while other models fail to reach the guessing entropy $1$ using as many as $5K$ traces, HierNet achieves the same with fewer than or close to $10$ traces.