On the Feasibility of Single-Trace Attacks on the Gaussian Sampler Using a CDT

We present a single-trace attack against lattice-based KEMs using the cumulative distribution table for Gaussian sampling and execute it in a real-world environment. Our analysis takes a single power trace of the decapsulation algorithm as input and exploits leakage of the Gaussian sampling subroutine to reveal the session key. We investigated the feasibility of the attack on different boards and proved that the power consumption traces become less informative with higher clock frequencies. Therefore, we introduce a machine-learning denoising technique, which enhances the accuracy of our attack and leverages its success rate to 100%. We accomplish the attack on FrodoKEM, a lattice-based KEM and third-round alternate candidate. We execute it on a Cortex-M4 board equipped with an STM32F4 micro-controller clocked at different frequencies.