August_EDFA_Digital

edfas.org ELECTRONIC DEV ICE FA I LURE ANALYSIS | VOLUME 24 NO . 3 12 SUPERVISED FEATURE EXTRACTION AND SYNTHESIS OF INTEGRATED CIRCUITS MICROGRAPHS FOR PHYSICAL ASSURANCE Md. Mahfuz Al Hasan, Md. Tahsin Mostafiz, and Navid Asadizanjani Florida Institute for Cybersecurity Research, University of Florida, Gainesville, Florida mdmahfuzalhasan@ufl.edu EDFAAO (2022) 3:12-22 1537-0755/$19.00 ©ASM International ® INTRODUCTION In a semiconductor supply chain, an untrusted foun- dry is an outsourced fabrication entity that holds access to circuit design information, such as layouts and test vectors. [1] Recently, image processing-based physical inspection techniques have drawn much attention be- cause of the advancement in the microscopy field as a means to detect malicious circuitry (such as logic cells) insertion, deletion, or modification which are known as hardware Trojans, [2] or cloned intellectual property (IP) in a counterfeit integrated circuit (IC) by such untrusted foundries. [3] Scanning electron microscopy (SEM) image analysis using artificial intelligence (AI) techniques such as machine/deep learning has established itself for this purpose. A classical machine learning-based discrimina- tive classifier, such as a support vector machine (SVM), needs pre-defined features that can be extracted from the cells of an SEM image for training. Nevertheless, the sys- temcomeswith amultitude of limitations including a high dependency on the pre-processing steps such as inten- sity equalization and segmentation. [4] In the presence of slight unwanted situations like image noise, out-of- distribution data, etc., this dependency leads to improper feature extraction and yields an unreliable trojan detec- tion system. Deep convolutional neural networks (CNN) have been successful in extracting meaningful features from natural images [5] and so directly operating them on the SEM images should enhance the generalization capability of the system. However, a comprehensive dataset is required to achieve the objective while using deep CNN, which is not always available. To build a dataset one cannot solely rely on the IC image acquisition process. Several specific parameters involved in this process include dwelling time, magnification, brightness, and contrast which can nega- tively impact theacquisition time. Inaddition, imaging varia- tions due to the quality of the sample polishing process fur- ther reduce the possibility of gathering meaningful data. This article proposes a design for a real-time Trojan detection system and explores the possibility of being independent of the large-scale SEM image acquisition process. A deep-learning approach has been adopted to generate synthetic micrographs i.e., full-scale synthetic SEM images from layout images to show that in the future, it’s possible to synthesize SEM images directly from layout images rather than depending on the time-consuming image acquisition process. Then, a computer vision (CV)- based approach capable of extracting cells available from the images has been applied (see Fig. 1). To combat the data problem, cell images have been synthesized from the extracted ones using a generative adversarial network (GAN). Finally, features from real and synthesized cells have been extracted and classified by a convolutional neural network (CNN) in a supervised manner. The clas- sification outcome is matched with the design exchange format file entry to ensure the purity of the underlying IC. These contributions are summarized as follows: • A layout to the SEM image translation unit to reduce the dependency on the image acquisition. • An automated extraction unit to extract cells from the IC SEM images. • A synthetic IC SEM image generation unit based on mode-seeking GAN (MSGAN) to synthesize diverse cell images. • A cell recognition unit to classify diverse cell types. This concludes every part of the end-to-end system.