Deep-learning side-channel analysis of embedded ML-KEM: Single-trace message recovery, cross-device generalization, and frequency-band leakage characterization

Quantum computing undermines traditional public-key cryptosystems like RSA and ECC, prompting NIST to standardize Module Lattice Key Encapsulation Mechanism (ML-KEM) as a core Post-Quantum Cryptography (PQC) scheme in August 2024. Despite its theoretical security, ML-KEM’s practical vulnerability in embedded systems remains underexplored. This study evaluates ML-KEM on an ARM Cortex-M4 microcontroller under deep learning(DL)-based side-channel attacks (DL-SCA), revealing that a single power trace enables 32-byte message recovery with 99.98% accuracy, while cross-device testing across three chips achieves over 99% accuracy, highlighting critical DL-driven threats. To better understand the spectral characteristics exploited by deep-learning-based side-channel attacks, we conduct a controlled frequency-band sensitivity study using Butterworth and Chebyshev low-pass filters applied to recorded traces. Across different cutoff frequencies and filter orders, the validation accuracy decreases substantially under aggressive attenuation, indicating that a significant fraction of exploitable leakage resides in higher-frequency components. Rather than claiming that offline digital filtering constrains a realistic adversary, we interpret these results as evidence for where leakage-sensitive information is concentrated, thereby providing guidance for implementable countermeasures such as masking, shuffling, desynchronization, and acquisition-path power/EM conditioning in embedded PQC systems. This cost-efficient solution requires no algorithm/hardware modifications, emphasizing the urgency of implementation-level protections for secure PQC deployment in resource-constrained systems.