CryptoDB
Adaptive Template Attacks on the Kyber Binomial Sampler
| Authors: | |
|---|---|
| Download: | |
| Abstract: | Template attacks build a Gaussian multivariate model of the side-channel leakage signal generated by each value of a targeted intermediate variable. Combined with additional steps, such as dimensionality reduction, such models can help to infer a value with nearly 100% accuracy from just a single attack trace. We demonstrate this here by reconstructing the output of the binomial sampler of a Cortex-M4 imple- mentation of the Kyber768 post-quantum key-encapsulation mechanism. However, this performance is usually significantly diminished if the device, or even just the ad- dress space, used for profiling differs from the attacked one. Here we introduce a new technique for adapting templates generated from profiling devices in order to attack another device where we are also able to record many traces, but without knowledge of the random value held by the targeted variable. We interpret the model from the profiling devices as a Gaussian mixture and use the Expectation–Maximization (EM) algorithm to adapt its means and covariances to better match the unlabelled leakage distribution observed from the attacked setting. The Kyber binomial sampler turned out to be a particularly suitable target, for two reasons. Firstly, it generates a long sequence of values drawn from a small set, limiting the number of Gaussian components that need to be adjusted. Secondly, the length of this sequence requires particularly well-adapted templates to achieve a high key-recovery success rate from a single trace. We also introduce an extended point-of-interest selection method to improve linear discriminant analysis (LDA). |
BibTeX
@article{tches-2025-35786,
title={Adaptive Template Attacks on the Kyber Binomial Sampler},
journal={IACR Transactions on Cryptographic Hardware and Embedded Systems},
publisher={Ruhr-Universität Bochum},
volume={2025},
pages={470-492},
url={https://tches.iacr.org/index.php/TCHES/article/view/12224},
doi={10.46586/tches.v2025.i3.470-492},
author={Eric Chun-Yu Peng and Markus G. Kuhn},
year=2025
}