Reimagining the Machine Vision Pipeline in Cyber Physical Systems for Trustworthiness and Efficiency
Date of Award
12-20-2024
Degree Name
Doctor of Philosophy (PhD)
Degree Type
Dissertation
Abstract
Cyber-physical systems (CPS), including autonomous vehicles, drones, and mobile robots, rely on intricate sensors, actuators, and machine learning algorithms to perceive the physical world and execute actions within their surroundings. In the context of vision-driven CPS, achieving this demands processing a substantial volume of visual data captured by on-board cameras. The data is subsequently channeled through digital processors and harnessed by deep neural networks for tasks such as image classification, object detection, and depth perception. This data-centric, machine-vision-infused CPS fosters intelligent decision-making, thereby enhancing overall system performance. Trustworthiness, encompassing the robustness of the entire machine-vision pipeline, and system-level efficiency are two primary challenges that vision-driven CPS must confront under stringent power and resource constraints, while simultaneously achieving low-latency decision-making to avert real-world repercussions. This dissertation addresses each of these characteristics by identifying vulnerabilities in existing state-of-the-art systems and redesigning the image acquisition process itself to improve the overall efficiency of machine vision enabled CPS. First, physically realizable vulnerabilities in autonomous driving vehicles are found by modifying the environment perceived by the camera. Then a camera lens is modified to validate a supply chain attack that allows an object to be hidden in plain sight. Diffractive optical neural networks are used to perform pre-capture privacy enhancement. A novel image sensor is implemented to perform task-specific compression during image capture. Finally, the trade-off between trustworthiness and efficiency is analyzed.
Language
English (en)
Chair
Yevgeniy Vorobeychik
Committee Members
Ayan Chakrabarti; Christopher Gill; Ning Zhang; Xuan Zhang