Curvature Consistent Network for Microscope Chip Image Super-Resolution.

Area of Science:

• Computer Vision
• Integrated Circuit Security
• Image Processing

Background:

• Hardware Trojan (HT) detection is critical for secure infrastructure.
• High-resolution (HR) microscope chip images (MCIs) are costly and time-consuming to acquire.
• Low-resolution (LR) MCIs offer faster, cheaper acquisition but suffer from lost details and noise, hindering HT detection.

Purpose of the Study:

• To propose a novel MCI super-resolution (SR) method to improve HT detection from LR images.
• To investigate the importance of recovering curvature information for accurate HT detection.
• To address the challenge of limited MCI datasets by creating a new benchmark.

Main Methods:

• A curvature consistent network (CCN) employing homogeneous and heterogeneous workflows.
• Homogeneous workflow: learns LR-HR MCI mapping.
• Heterogeneous workflow: learns MCI-curvature image mapping.
• Collaborative fusion strategy to combine features from both workflows for HR image recovery.

Main Results:

• The proposed CCN method outperforms existing SR methods in recovering delicate circuit lines.
• The CCN method significantly improves HT detection performance on LR MCIs.
• A new benchmark dataset (MCI) of realistic MCIs at various resolutions was created and used for evaluation.

Conclusions:

• Recovering curvature information is vital for enhancing HT detection from MCIs.
• The CCN method effectively addresses the limitations of LR MCIs for HT detection.
• The developed MCI dataset and CCN method provide valuable resources for future research in hardware security and image super-resolution.