Introduction
As integrated circuits (ICs) shrink to the nanoscale, they become increasingly vulnerable to "Hardware Trojans"—malicious modifications hidden within billions of transistors. Ensuring hardware assurance, or verifying that a chip matches its design, currently relies on Scanning Electron Microscopy (SEM) and Reverse Engineering (RE). However, this process is a slow, expensive bottleneck, producing images riddled with noise and artifacts. While computer vision (CV) has addressed similar challenges in medical imaging, the hardware assurance space has suffered from a "data desert."
The REFICS project addresses this by providing a high-fidelity synthetic alternative to costly physical imaging, bridging the gap between CV research and hardware security.
Methodology: Synthetic SEM Generation
The core contribution is a sophisticated synthetic SEM image generation workflow. Rather than applying simple noise to a layout, the authors simulate the physics of electron-material interaction. This includes modeling Poisson shot noise from the scanning beam, Gaussian noise from material response, and real-world hardware issues like beam drift and sensor noise.
The workflow consists of several stages:
- Beam Interactions: Modeling noise at the source.
- Predictable & Random Noise: Simulating hardware artifacts like "stitching errors" where image patches misalign.
- Operator Interactions: Including variations in contrast and brightness typical of manual SEM operation.
Validation using Jensen-Shannon divergence and Fourier domain analysis proved these synthetic images are statistically nearly identical to real SEM captures.
The REFICS Dataset
The REFICS dataset comprises 800,000 synthetic SEM images covering four cardinal IC layers (doping, polysilicon, contacts, and metal) across two manufacturing nodes (32/28nm and 90nm). Each image includes a segmented ground truth and a layout-level mask. Unlike datasets for Printed Circuit Boards (PCBs), REFICS focuses on the significantly more complex internal layers of ICs.
Key Insights and Benchmarking
Benchmarking revealed a significant "Generalization Gap." Models trained on high-contrast layers like "Metal" fail when applied to low-contrast, complex layers like "Polysilicon." This suggests that off-the-shelf CV models lack the robustness required for cross-node hardware verification.
Furthermore, the authors identified a "Metric Mismatch." Standard vision metrics like SSIM or IoU do not always correlate with hardware integrity. A high IoU score might mask a tiny "open circuit" that renders a chip non-functional.
Key results from benchmarking include:
- Denoising: CBDNet performed best among deep learning models, while traditional Anisotropic Diffusion filters were surprisingly effective at preserving sharp edges.
- Segmentation: Supervised models outperformed unsupervised ones but lacked cross-layer and cross-node generalizability. Performance dropped sharply when a model trained on 32nm chips was tested on 90nm chips.
Conclusion and Future Work
REFICS provides a critical benchmark for the hardware assurance community. However, synthetic data remains a proxy. Future efforts will focus on releasing real-world SEM sub-datasets, developing connectivity-aware metrics that prioritize electrical continuity over pixel accuracy, and expanding from 2D patches to 3D alignment for full-chip verification.
Summary updated