Semantic-Masked Intensity Augmentation for Deep Learning-based Analysis of FPGA Images

The emergence of data science and deep learning has enabled the automated recognition of circuit elements from the microscopic images of delayered Integrated Circuit (IC) devices, and has greatly improved the efficiency of overall functional analysis flow for hardware security. However, due to the high complexity of delayering the manufactured IC devices and the imaging imperfections in modern ICs, the acquired microscopic images usually contain unforeseeable variations even for the same types of circuit elements. As a result, the deep learning model which is typically trained with a very limited set of labelled images suffers from inefficacy on generalizing to unseen images, which further causes errors for subsequent analysis. Data augmentation techniques, which virtually introduce data variations and increase the data amount by applying different image transformations, are thus widely used during the training of deep learning models for IC image analysis. In this paper, we propose a Semantic-Masked Intensity Augmentation (SMIA) technique with a deep-learning-based framework to analyze the microscopic images acquired from a delayered Field-Programmable Gate Arrays (FPGA) device. Different from the commonly-used intensity augmentation techniques which apply transformations to the image pixels according to their original intensities, our proposed SMIA considers the semantic context of the image pixels by applying different intensity transformations according to pixel-level semantic masks. With experiments on segmenting metal lines from the metal layer images of a targeted FPGA, our proposed SMIA demonstrates better performance and higher stability than the existing intensity augmentation techniques.