International Association
for Cryptologic Research

Deep learning-based side-channel analysis (DLSCA) has demonstrated remarkable performance over the past few years. Even with limited preprocessing and feature engineering, DLSCA is capable of breaking protected targets, sometimes requiring only a single attack trace. In the DLSCA context, the commonly investigated countermeasures are Boolean masking and desynchronization. While the exact mechanisms of how DLSCA breaks masking are less understood, the core idea behind handling desynchronization is simple. Convolutional neural networks (CNNs) are shift invariant, allowing them to overcome desynchronization. However, considering the importance and practicality of desynchronization countermeasures, we know remarkably little about the limits of CNNs or how to enhance their capabilities when dealing with desynchronization.

In this work, we begin with the theoretical foundations of shift and temporal scale equivariance. Afterward, we build a neural network model allowing such equivariance and test it against several commonly considered targets. Our results demonstrate that equivariant CNNs are robust, easy to design, and achieve excellent attack performance. More precisely, we showcase how such a simple model can even outperform recent transformer-based neural networks. Finally, we demonstrate the practical relevance of scale equivariance by showing how an equivariant CNN can learn leakage from a device operating at one clock frequency and generalize to a device with a different clock frequency, a result not previously demonstrated in DLSCA.