Date of Award

Spring 2022

Document Type

Dissertation

Degree Name

Doctor of Engineering (DEng)

Department

Electrical Engineering (ENAS)

First Advisor

Szefer, Jakub

Abstract

As computer systems grow more and more complicated, various optimizations can unintentionally introduce security vulnerabilities in these systems. The vulnerabilities can lead to user information and data being compromised or stolen. In particular, the ending of both Moore's law and Dennard scaling motivate the design of more exotic microarchitectural optimizations to extract more performance -- further exacerbating the security vulnerabilities. The performance optimizations often focus on sharing or re-using of hardware components within a processor, between different users or programs. Because of the sharing of the hardware, unintentional information leakage channels, through the shared components, can be created. Microarchitectural attacks, such as the high-profile Spectre and Meltdown attacks or the cache covert channels that they leverage, have demonstrated major vulnerabilities of modern computer architectures due to the microarchitectural~optimizations. Key components of processor microarchitectures are processor caches used for achieving high memory bandwidth and low latency for frequently accessed data. With frequently accessed data being brought and stored in caches, memory latency can be significantly reduced when data is fetched from the cache, as opposed to being fetched from the main memory. With limited processor chip area, however, the cache size cannot be very large. Thus, modern processors adopt a cache hierarchy with multiple levels of caches, where the cache close to processor is faster but smaller, and the cache far from processor is slower but larger. This leads to a fundamental property of modern processors: {\em the latency of accessing data in different cache levels and in main memory is different}. As a result, the timing of memory operations when fetching data from different cache levels, e.g., the timing of fetching data from closest-to-processor L1 cache vs. from main memory, can reveal secret-dependent information if attacker is able to observe the timing of these accesses and correlate them to the operation of the victim's code. Further, due to limited size of the caches, memory accesses by a victim may displace attacker's data from the cache, and with knowledge, or reverse-engineering, of the cache architecture, the attacker can learn some information about victim's data based on the modifications to the state of the cache -- which can be observed by the timing~measurements. Caches are not only structures in the processor that can suffer from security vulnerabilities. As an essential mechanism to achieving high performance, cache-like structures are used pervasively in various processor components, such as the translation lookaside buffer (TLB) and processor frontend. Consequently, the vulnerabilities due to timing differences of accessing data in caches or cache-like structures affect many components of the~processor. The main goal of this dissertation is the {\em design of high performance and secure computer architectures}. Since the sophisticated hardware components such as caches, TLBs, value predictors, and processor frontend are critical to ensure high performance, realizing this goal requires developing fundamental techniques to guarantee security in the presence of timing differences of different processor operations. Furthermore, effective defence mechanisms can be only developed after developing a formal and systematic understanding of all the possible attacks that timing side-channels can lead to. To realize the research goals, the main main contributions of this dissertation~are: \begin{itemize}[noitemsep] \item Design and evaluation of a novel three-step cache timing model to understand theoretical vulnerabilities in caches \item Development of a benchmark suite that can test if processor caches or secure cache designs are vulnerable to certain theoretical vulnerabilities. \item Development of a timing vulnerability model to test TLBs and design of hardware defenses for the TLBs to address newly found vulnerabilities. \item Analysis of value predictor attacks and design of defenses for value predictors. \item Evaluation of vulnerabilities in processor frontends based on timing differences in the operation of the frontends. \item Development of a design-time security verification framework for secure processor architectures, using information flow tracking methods. \end{itemize} \newpage This dissertation combines the theoretical modeling and practical benchmarking analysis to help evaluate susceptibility of different architectures and microarchitectures to timing attacks on caches, TLBs, value predictors and processor frontend. Although cache timing side-channel attacks have been studied for more than a decade, there is no evidence that the previously-known attacks exhaustively cover all possible attacks. One of the initial research directions covered by this dissertation was to develop a model for cache timing attacks, which can help lead towards discovering all possible cache timing attacks. The proposed three-step cache timing vulnerability model provides a means to enumerate all possible interactions between the victim and attacker who are sharing a cache-like structure, producing the complete set of theoretical timing vulnerabilities. This dissertation also covers new theoretical cache timing attacks that are unknown prior to being found by the model. To make the advances in security not only theoretical, this dissertation also covers design of a benchmarking suite that runs on commodity processors and helps evaluate their cache's susceptibility to attacks, as well as can run on simulators to test potential or future cache designs. As the dissertation later demonstrates, the three-step timing vulnerability model can be naturally applied to any cache-like structures such as TLBs, and the dissertation encompasses a three-step model for TLBs, uncovering of theoretical new TLB attacks, and proposals for defenses. Building on success of analyzing caches and TLBs for new timing attacks, this dissertation then discusses follow-on research on evaluation and uncovering of new timing vulnerabilities in processor frontends. Since security analysis should be applied not just to existing processor microarchitectural features, the dissertation further analyzes possible future features such as value predictors. Although not currently in use, value predictors are actively being researched and proposed for addition into future microarchitectures. This dissertation shows, however, that they are vulnerable to attacks. Lastly, based on findings of the security issues with existing and proposed processor features, this dissertation explores how to better design secure processors from ground up, and presents a design-time security verification framework for secure processor architectures, using information flow tracking methods.

COinS