UC Irvine

Since the appearance of microprocessors, miscellaneous categories of computer systems, including Embedded and Cyber-Physical Systems (ECPS), have become an integral part of our modern society. In contrast to general-purpose computers, an ECPS is a computing system designed to perform a dedicated or narrow range of functions with minimal human intervention. Today, the advanced standards in device networking, ubiquitous access to the Internet, miniaturization of processors, and reduced power consumption have taken this field to the next level. However, despite these technological growths, designing a robust ECPS remains an open research challenge where the goal is to achieve better functionalities, encompass complex, uncertain, and changing environments, and ensure system security. Minor failures on ECPS may cause severe collapses or cyberattacks, impeding progress toward increasing automation and modernizing our computing environments.

To achieve robustness, this dissertation studies the embedding of Computational Intelligence (CI) into emerging Graph Learning (GL) technologies. The CI paradigms mimic the nature of humans, aiming at solving complex problems and exhibiting a cognitive ability to learn or adapt to new situations to generalize, abstract, discover, and associate. As a prerequisite of this research direction, Machine Learning (ML) has become increasingly ubiquitous, with existing works exploring various fields, from self-driving vehicles, facial recognition systems, and real-time language translation to security surveillance, innovative home applications, and health monitoring. However, conventional ML algorithms typically require appropriate vectorized representations crafted by domain experts to accomplish the desired goals. Graph-structured data have imposed unprecedented challenges on ML due to their inherent complexity. Unlike text, audio, and images, graphs are embedded in an irregular dimension, making some essential operations of ML inapplicable. GL has attracted much attention to new research ideas in several fields. To date, many researchers have proven the usefulness of GL in social computing, information retrieval, computer vision, bioinformatics, economics, and e-commerce. However, its applications in the subfields of ECPS still need to be explored.

In this dissertation, we will cover GL applications for ECPS in robust Integrated Circuit (IC) design analysis for hardware security, robust binary analysis for enabling software security, and ultimately how GL brings the scene-understanding capabilities of autonomous driving systems to the next level. In IC design analysis (Chapter 2), we explore how GL can be leveraged to resolve challenging problems in hardware security. We propose HW2VEC and demonstrate how it can be utilized for Hardware Trojan (HT) detection and Intellectual Property (IP) piracy detection. Next, in the binary analysis (Chapter 3), we demonstrate how a more advanced GL methodology, called CFG2VEC, can reverse-engineer the semantic source-level information lost during the compilation process, thus making the binary software patching tasks more efficient. Then, we describe our novel methodology, SG2VEC, to enhance autonomous driving systems' scene understanding capabilities (Chapter 4). To enable this cognitive capability, we propose to use scene-graph to encode the surrounding traffic participants and a pipeline of spatiotemporal scene-graph embedding networks to process scene-graphs and learn toward goals.