Automated Stitching for Scanning Electron Microscopy Images of Integrated Circuits

Abstract

While established toolchains are widely available for reverse engineering software, firmware, and even hardware on the printed circuit board level to some degree, the entry barrier to reverse engineering of integrated circuits (ICs) remains high due to the associated cost of equipment. One key driver of this high cost is the requirement of high-quality scanning electron microscopes (SEMs) for the analysis of ICs with small feature sizes. While optical microscopes are a cost-effective alternative for large feature sizes, modern ICs are manufactured with feature sizes that are too small for the limited magnification of optical microscopes. In IC analysis, SEMs are used to create images of individual layers of integrated circuits during an iterative delayering process. The aim of this process is to image logic gate placement and interconnects, from which detailed information about the implemented logic functions can be recovered. In some cases, it is also possible to extract the contents of read-only memory (ROM) on the chip. An SEM usually creates images in the range of megapixel resolutions, but analyzing an IC layer requires resolutions in the gigapixel range. To create such large images, many individual images must be taken and then fused into one large image. Compared to images created by optical microscopes, SEM images pose unique challenges: They are affected by distortion due to charging effects and often exhibit high levels of noise and low contrast. One way of reducing the entry barrier to IC reverse engineering is to develop algorithms that can provide good results even in the case of suboptimal image quality, as can be produced by comparatively cheap, used SEMs. This thesis introduces and evaluates several algorithms for the purpose of fusing noisy images with low contrast created by older SEMs. Based on an evaluation using gigapixel scale image sets, the most efficient and effective image registration algorithm for these image properties is determined. These algorithms determine offsets between individual overlapping images in the image set. For the two best algorithms, automated inference of optimal parameters is developed. Four global stitching algorithms are introduced, to create large fused images based on the results of image registration. These four algorithms optimize for different quality metrics in the generated fused image. Finally, the introduced algorithms are evaluated and compared to state-of-the-art image stitching software.

Type
Publication
TU Wien