August_EDFA_Digital

edfas.org 13 ELECTRONIC DEV ICE FA I LURE ANALYSIS | VOLUME 24 NO . 3 TROJAN DETECTION TECHNIQUES Over the last decade, there’s beenan intensive focus on developing techniques to detect hardware Trojans using physical inspection techniques due to the availability of high accuracy polishingmethods, easy access to advanced nanoimaging tools on a rent basis, and application devel- opment in image analysis techniques. Initially, Courbon et al. [6] proposed a very basic concept of using image subtraction between the IC’s frontside images of the IC under inspection and a golden IC to detect Trojans; it was a simple idea involving just a single SEM image, but it lacks any hardware Trojan attack analysis and detection in the presence of imaging or process variations. Vashistha et al. [4] also used a single image-based correlated technique using a low-quality SEM image of the silicon die and com- pared it with a high-quality SEM image to mimic a golden layout. This technique uses a structural similarity index to determine the existence of Trojans at the active layer. Bao et al. [7] proposed a feature-based technique for Trojans at metal layers usingmachine learning classifiers. Thismethoddoes not pay attention toTrojans implement- ed at the active layer. Also, their performance evaluation was conducted on synthesized images with only additive Gaussian noise. Since real SEM images of IC layers contain process variations and imaging noises, theirmethod could easily fail in practical applications. Vashistha et al. [3] proposed the Trojan Scanner (TS) method that compares transistors’ active layer footprints with the golden layout to detect malicious changes. This framework used a machine learning-based framework to detect malicious changes. Since the layout was used as a golden standard, it can result in false positives detec- tion because the layout features lack variation due to the fabrication process and SEM imaging. All the techniques mentioned earlier used standalone conventional image analysis. Hence, they failed tocapture variations in images, extract detailed features from images, and lead to poor confidence in hardware Trojan detection. To face those challenges, theproposedmethod focuses on synthesizing SEM cell images and presenting an end- to-end Trojan Detection System through the analysis of SEM images. In addition to that, to remove the high dependency on SEM image acquisition, a state-of-the-art layout to SEM image synthesis/translation technique is discussed. To reduce the bottleneck for image acquisi- tion time, SEM images are synthesized from the layout of respective IC in an adversarial manner. Then using the traditional image processing technique, cells were extracted from SEM images. To handle the data-scarcity problem, cell images were synthesized using the real ones. Then using real and synthesized cell images, a convolutional neural network (CNN) classifier is trained. During inference, cells are extracted one-by-one row-wise from the SEM image of IC under authentication. Each of the cells is passed through the trained classifier and the output of the classifier is matched with the DEF entry of the IC. Based on the outcome, a Trojan presence can be detected. A schematic diagram of the system during the inference stage has been mentioned in Fig. 1. Each step of the proposed method along with experimental results has been discussed in the following literature. LAYOUT TO HIGH-RESOLUTION SEM IMAGE TRANSLATION Todevelop an end-to-end image analysis-basedTrojan detection system, the availability of a diverse set of data is a requisite. To mitigate the time-cost of SEM image acquisition, high-res (1024 x 1024) SEM images have been synthesized from the layout of the chip. This essentially paves the way for generating an SEM-like image from the Fig. 1 An end-to-end scheme for Trojan detection system. Cells are localized and passed one-by-one through the trained CNN classifier. The classifier outputs a probability distribution over the number of classes. For each position of the IC, the classifier output is matched with the corresponding DEF file entry and based on the agreement a Trojan presence is decided. The classifier is trained using localized and synthetic cells generated from real ones.