Overcoming Data Scarcity and Confidentiality in Hardware Assurance via Synthetic Generation

{ “meta”: “This paper presents a synthetic data generation pipeline for hardware assurance that addresses data scarcity and confidentiality issues.”, “body”: “## Problem\nThis work addresses the challenge of data scarcity and confidentiality in hardware assurance, particularly in the context of scanning electron microscopy (SEM) for verifying nanoscale structures. The authors highlight the difficulties in assembling large, high-quality datasets due to the time-intensive nature of data acquisition and strict intellectual property (IP) constraints on proprietary designs. Notably, this is a preprint and has not undergone peer review.\n\n## Method\nThe proposed pipeline employs a two-step generative approach to create a synthetic dataset while preserving IP confidentiality. First, a StyleGAN is utilized to learn the distribution of hardware layout masks, generating novel structures that exhibit macroscopic variation. Following this, a conditional GAN (Pix2PixHD) translates these masks into realistic SEM images, ensuring that the generated images maintain authentic textures and noise characteristics. The segmentation model is then trained exclusively on this synthetic dataset, enabling a successful transfer to real images.\n\n## Results\nThe available text does not report quantitative results. However, it is stated that the segmentation model trained on synthetic data outperforms a baseline model that was trained on a limited real dataset, demonstrating effective “