Privacy-Preserving Synthetic SEM Dataset Generation for Hardware Assurance via GANs

Forum topic · 小凯 · 2026-08-12

Summary

Hardware assurance based on scanning electron microscopy (SEM) depends on large, high-quality datasets, but assembling them is difficult because acquisition is time-consuming and proprietary designs are protected by strict IP constraints. This paper proposes a privacy-preserving pipeline that safeguards IP by heavily distorting functional layouts while producing a visually realistic synthetic dataset from only a small set of initial samples. A StyleGAN first learns the distribution of hardware layout masks to generate novel, macroscopically diverse structures. A conditional GAN (Pix2PixHD) then translates these masks into realistic SEM-style images that preserve authentic textures and noise. The authors train segmentation models exclusively on the synthetic data and show successful sim-to-real transfer, with their models outperforming baselines trained on the limited real dataset. Because the synthetic layouts are novel and do not replicate any proprietary routing, downstream segmentation models reduce exposure to gradient inversion and membership inference attacks, yielding a high-performance, security-conscious solution for nanoscale hardware verification.

Overview

Field: Computer Vision (CV)
Authors: Gijung Lee, Ronald Wilson, Damon L. Woodard
Published: 2026-08-12
arXiv: 2508.05150

Background

Hardware assurance verifies nanoscale structures using scanning electron microscopy (SEM). Automated analysis requires large, high-quality datasets, but collecting them is hard:

Acquisition is slow and time-intensive.
Proprietary designs are bound by strict intellectual property (IP) constraints.

These two obstacles motivate a method that can train useful models without exposing sensitive layout data.

Proposed Pipeline

A two-stage privacy-preserving pipeline generates a synthetic dataset from only a small number of initial samples.

1. StyleGAN — mask synthesis

Learns the distribution of hardware layout masks.
Generates novel, macroscopically varied structures.

Pix2PixHD (conditional GAN) — image synthesis

Translates the generated masks into realistic SEM-style images.
Preserves authentic textures and noise characteristics.

Crucially, the initial functional designs are heavily distorted. The resulting synthetic layouts are novel structures and do not replicate any specific proprietary routing of the source designs.

Key Findings

Segmentation models trained exclusively on the synthetic data transfer successfully to real SEM images (sim-to-real transfer).
These synthetic-only models outperform baseline segmentation models trained on the limited real dataset.
Because the underlying synthetic layouts do not copy specific proprietary wiring, deploying the final segmentation pipeline reduces the risk of exposing sensitive IP to:
Gradient inversion attacks
Membership inference attacks

Significance

The paper delivers a high-performance, security-conscious solution for nanoscale hardware verification: it removes the need to share or annotate large volumes of proprietary SEM imagery, while still enabling accurate downstream automated analysis.

Reference

arXiv: https://arxiv.org/abs/2508.05150

--- *Auto-collected 2026-08-12*