Generative AI for High-Fidelity Side-Channel Trace Simulation

This research began through the NTU Global Connect Fellowship Program at Nanyang Technological University and continues as a remote collaboration, supervised by Gwee Bah Hwee and Juncheng Chen. A manuscript is in preparation.

The project studies how attack-relevant leakage emerges during diffusion-based side-channel trace generation and whether cryptanalytic utility develops at the same rate as waveform and distribution fidelity.

Experiments and Findings

• Datasets: Developed conditional DDPM experiments on controlled synthetic traces and four measured AES datasets.
• Reverse-trajectory evaluation: Evaluated predicted-clean traces across all 400 reverse-diffusion stages rather than only the final samples.
• Attack-based evaluation: Measured CPA-based key-recovery complexity throughout the reverse trajectory.
• Fidelity evaluation: Combined waveform and distribution distances with nearest-neighbor likelihood scoring.
• Key observation: Found that attack-relevant leakage can emerge substantially earlier than full distributional fidelity.
• Masked traces: Matching first-order marginals alone was insufficient to reproduce the cross-share dependence required for second-order attacks.