Back to search

University of Washington

LSTAR Framework: Lightweight Framework for Standardizing Tests for Adversarial Robustness

Abstract

dc:description.abstract

The role of neural networks in various tasks has exploded in recent years, becoming prevalentin many safety-critical applications. However, improving neural network robustness has be- come a challenge due to the existence of adversarial examples—imperceptible perturbations to the inputs of machine learning models that mislead classifiers into producing incorrect outputs. While there have been numerous advancements in crafting adversarial attacks and defenses, research on the basis of adversarial examples has notably lagged behind, largely due to the computational difficulty of analyzing high-dimensional spaces. This inherent difficulty has led researchers to construct models for understanding adversarial examples divergent from conventional paradigms, with some relying on commonly used frameworks while others utilize their own tailored frameworks to meet their unique needs. Consequently, replicating and building upon research in this field presents a significant challenge. In this paper, we present a modular, lightweight framework to assist researchers in ad-dressing these challenges by providing a comprehensive approach to evaluating machine learning models through a standardized experimentation platform. We present several po- tential hypotheses regarding the basis of adversarial examples and utilize our framework to verify them more robustly under complex attacks and datasets through controlled experi- ments. Our experimental results indicate that geometric causes directly affect the robustness of machine learning models, while statistical factors amplify the effects of adversarial attacks. This framework provides a baseline for further studies to better understand the phenomenon of adversarial examples, allowing researchers to design more robust machine learning models.

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Tran, Kenneth
Advisor dc:contributor.advisor
  • Lagesse, Brent

Subjects

dc:subject × 8

Rights

dc:rights
Statement dc:rights
  • CC BY
Language dc:language.iso
en_US

Identifiers

dc:identifier.*
Handle dc:identifier.uri
https://hdl.handle.net/1773/51652

Chain of custody

source
Harvested from
University of Washington
Base URL
digital.lib.washington.edu/server/oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Tran, Kenneth. LSTAR Framework: Lightweight Framework for Standardizing Tests for Adversarial Robustness. 2024. https://hdl.handle.net/1773/51652