Date of Award

Fall 2023

Project Type

Dissertation

Program or Major

Electrical and Computer Engineering

Degree Name

Doctor of Philosophy

First Advisor

Qiaoyan Yu

Second Advisor

LaCourse LaCourse

Third Advisor

Shaad Md Mahmud

Abstract

Approximate Computing (AC) techniques have been leveraged to improve computing performance and energy efficiency with minor degradation in accuracy. Different than conventional computing, AC allows the computation to deviate from the reference or deterministic execution behavior. Thus, AC has emerged as a new paradigm of computing systems, especially for the applications such as image processing, audio recognition, and artificial intelligence. The majority of research efforts on approximate computing focus on new approximation mechanism design and implementation, rather than examining the security vulnerabilities of AC systems. To fill this gap, we first analyze the potential security threats induced by the AC techniques applied at various levels of the computing stack. Qualitative and quantitative analyses are performed to assess the impact of the new security threats. Next, we propose systematic attack models, which systematically cover the attacks that build covert channels, compensate errors, terminate error resilience mechanisms, and propagate errors. To thwart those projected attacks, we further develop a guideline for countermeasure designs, in which randomization, access prohibition, design obfuscation, and noise injection are the four main principles. Two detailed obfuscation schemes are developed to obscure the entrance of attack surfaces and eliminate the explicit transition between approximate and precise modes, thus improving AC systems' resilience against attacks that tamper with approximate precise boundaries. We further developed a detection mechanism that leverages the intermediate node signals to differentiate the approximation errors from the intentional fault injection. Our countermeasures have the potential to effectively assure AC systems obtain high availability, confidentiality, integrity, and anti-piracy capability.

Download

Share

COinS