High-Fidelity Conditional Side-Channel Trace Synthesis Using Diffusion Models
IEEE Sensors Journal · 15.07.2026 · 26 (14) · 21527-21540
DOI: 10.1109/jsen.2026.3701075 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Sensors Journal, erschienen am 15.07.2026. Verfasst von Zekai Zhang und 7 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 24 Literaturangaben. Gefördert von: Innovation and Technology Fund, CityUHK Internal, National Natural Science Foundation of China.
Abstract (Originalsprache)
The evaluation of hardware security through physical side-channel measurements heavily relies on the volume and diversity of acquired sensor data, encompassing sufficient time-series instances for varying operational states. Synthesizing high-fidelity artificial measurement signals offers a compelling solution to augment limited empirical sensor data; however, manual crafting is infeasible, and existing automated generative approaches often fail to capture complex physical noise profiles. Recently, denoising diffusion probabilistic models (DDPMs) have emerged as superior alternatives to generative adversarial networks (GANs) for modeling realistic, noisy data distributions. This article explores the application of DDPMs for synthesizing side-channel sensor traces. We propose a novel conditional diffusion framework utilizing a specialized U-Net architecture, equipped with residual blocks and self-attention mechanisms, to learn and replicate complex temporal measurement dynamics. Trained on preprocessed measurement traces acquired via the industry-standard ChipWhisperer sensing platform, our model enables precise, conditional signal generation corresponding to specific hardware states. A critical advantage of our approach is its remarkable data efficiency: the proposed model requires merely 2560 empirical measurement traces for training, highlighting the practical data efficiency of the proposed framework; however, such cross-study comparisons should be interpreted with caution when the experimental settings are not identical. Experimental results demonstrate the high physical fidelity of the synthesized signals. A classifier trained on empirical measurements achieves 97.7% label consistency accuracy when evaluated on our synthetic traces, closely rivaling the 99.6% SOTA accuracy and providing strong evidence that critical physical leakage characteristics are preserved under the evaluated setting. Furthermore, in hardware vulnerability assessments, analytical models augmented with our synthetic data achieve a guessing entropy (GE) of one, matching best-in-class performance. In simulated evaluation scenarios using these ChipWhisperer-based traces, models trained with our synthesized physical signals achieve 91.1% accuracy in same-device characterizations and 82.4% in challenging cross-device assessments. Ultimately, these generated traces serve as a promising signal augmentation tool, significantly enhancing profiling models and providing a more rigorous evaluation of hardware security under constrained sensor data conditions.
Förderung
- Innovation and Technology Fund
- CityUHK Internal
- National Natural Science Foundation of China
- Jiangsu Province 100 Foreign Experts Introduction Plan
- Scientific Research Innovation Capability Support Project for Young Faculty
- Guangdong Provincial Key Laboratory of Interdisciplinary Research and Application for Data Science
- Guangdong and Hong Kong Universities “1+1+1” Joint Research Collaboration Scheme
- Guangdong Basic and Applied Basic Research Foundation
Zitieren
APA
Zhang, Z., Chen, D., Dai, W., Hong, J., Chan, Y. H., Koç, Ç. K., Hung, P. S. Y., & Cheung, R. C. C. (2026). High-Fidelity Conditional Side-Channel Trace Synthesis Using Diffusion Models. IEEE Sensors Journal, 26(14), 21527-21540. https://doi.org/10.1109/jsen.2026.3701075
Vancouver
Zhang Z, Chen D, Dai W, Hong J, Chan YH, Koç ÇK, et al. High-Fidelity Conditional Side-Channel Trace Synthesis Using Diffusion Models. IEEE Sensors Journal. 2026;26(14):21527-21540. doi:10.1109/jsen.2026.3701075
BibTeX
@article{Zhang2026,
author = {Zhang, Zekai and Chen, Donglong and Dai, Wangchen and Hong, Jinfa and Chan, Yu Hin and Koç, Çetin Kaya and Hung, Patrick S. Y. and Cheung, Ray C. C.},
title = {High-Fidelity Conditional Side-Channel Trace Synthesis Using Diffusion Models},
journal = {IEEE Sensors Journal},
year = {2026},
volume = {26},
number = {14},
pages = {21527--21540},
doi = {10.1109/jsen.2026.3701075}
}
EUVIMED