Towards security analysis and design of confidential computing systems

• Main Author: Lou, Xiaoxuan
• Other Authors: Zhang Tianwei
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science
• Online Access: https://hdl.handle.net/10356/180639
• Institution: Nanyang Technological University

Confidential computing establishes an encrypted Trusted Execution Environment to ensure confidentiality and integrity protection. However, it still suffers from two significant problems: (1) Vulnerability to micro-architectural side-channel attacks; (2) Security and efficiency issues when handling emerging applications. For the first problem, I conduct a comprehensive security analysis of existing confidential computing systems. An end-to-end attack method named NASPY is proposed to reveal novel Neural Architecture Search models from the encrypted TEE black box. Then Aegis is proposed as a unified defense framework for mitigating confidential virtual machines from Hardware Performance Counter side channels. For the second problem, I design more novel confidential computing systems integrating with recently emerging workloads. A watermarking scheme is designed for verifying the ownership of deep learning models within the TEE sandbox. Furthermore, I also integrate confidential computing with serverless computing to design a novel fast-launched confidential serverless computing system Neuralyzer