Back to search

Cornell University

An Information-Theoretic Approach To Practical Side Channel Protection

Abstract

dc:description.abstract

Side channel attacks are quickly becoming a serious problem in many computational security contexts. Presently, side channel countermeasures are primarily developed per case study, and the degree of mathematical rigor varies across studies. In order to address this methodological gap in the study of side channels, we use information theory-driven mathematical models to form a more rigorous foundation for the study of common classes of side channels, including one-shot and replayable channels. Due to the broad selection of competing side channel leakage metrics in the literature, we take both an information-theoretic and an empirical approach to analyzing the usefulness and use cases of the most common metrics, including the recently proposed maximal leakage metric. Using maximal leakage as an operationally interpretable upper bound on side channel leakage, we show that linear combinations of no more than two deterministic protection schemes against one-shot attacks are optimal, which is a surprising result. Conventional approaches to protecting against side channels that add randomized noise turn out to be suboptimal. In addition to one-shot attacks, we also study the class of replayable attacks in which the adversary can test the protection scheme many times with the same input. Here, we develop theoretical information bounds governing how the total leakage increases with the number of replays. The total leakage must naturally be upper bounded, but existing information-theoretic work provides only unbounded characterizations of how the total leakage increases with the number of replays. We prove that the total leakage approaches its upper limit exponentially fast, and the exponent can be relatively easily computed given the protection scheme. Finally, we study inference-phase attacks on dynamically pruned convolutional neural networks (CNNs), utilizing our theoretical results. Dynamic pruning is an increasingly popular technique for optimizing the execution time of the inference phase. We show that these techniques create a highly informative side channel that can be used to perform one-input-one-secret attacks, given indirect access to the pruning decisions. Moreover, we show that protecting this channel without losing the execution time savings from the pruning is naturally difficult because of the characteristics of the channel. Utilizing our studies of the one-shot and replayable attacks, we devise protection schemes that can significantly reduce the leakage from these attacks.

Degree

thesis:*
Name thesis:degree_name
Ph. D., Electrical and Computer Engineering
Level thesis:degree_level
Doctor of Philosophy
Discipline thesis:degree_discipline
Electrical and Computer Engineering
Grantor
Cornell University
Year dc:date.issued
2021

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Wu, Benjamin Huang
Chair dc:contributor.chair
  • Suh, Edward
Committee members dc:contributor.committeemember
  • Wagner, Aaron B.
  • Shi, Runting

Subjects

dc:subject × 4

Rights

Language dc:language.iso
en

Identifiers

dc:identifier.*
Dc Identifier Bibid
15312766
Dc Identifier Other
ProQuest Submission ID: 12830
ProQuest Publication ID: 28775896

Chain of custody

source
Harvested from
Cornell University
Base URL
ecommons.cornell.edu/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Wu, Benjamin Huang. An Information-Theoretic Approach To Practical Side Channel Protection. Doctor of Philosophy thesis, Cornell University, 2021. https://hdl.handle.net/1813/110906