AutoAdvExBench: Benchmarking autonomous exploitation of adversarial example defenses

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Carlini, Nicholas, Rando, Javier, Debenedetti, Edoardo, Nasr, Milad, Tramèr, Florian
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910855898071040
author Carlini, Nicholas
Rando, Javier
Debenedetti, Edoardo
Nasr, Milad
Tramèr, Florian
author_facet Carlini, Nicholas
Rando, Javier
Debenedetti, Edoardo
Nasr, Milad
Tramèr, Florian
contents We introduce AutoAdvExBench, a benchmark to evaluate if large language models (LLMs) can autonomously exploit defenses to adversarial examples. Unlike existing security benchmarks that often serve as proxies for real-world tasks, bench directly measures LLMs' success on tasks regularly performed by machine learning security experts. This approach offers a significant advantage: if a LLM could solve the challenges presented in bench, it would immediately present practical utility for adversarial machine learning researchers. We then design a strong agent that is capable of breaking 75% of CTF-like ("homework exercise") adversarial example defenses. However, we show that this agent is only able to succeed on 13% of the real-world defenses in our benchmark, indicating the large gap between difficulty in attacking "real" code, and CTF-like code. In contrast, a stronger LLM that can attack 21% of real defenses only succeeds on 54% of CTF-like defenses. We make this benchmark available at https://github.com/ethz-spylab/AutoAdvExBench.
format Preprint
id arxiv_https___arxiv_org_abs_2503_01811
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle AutoAdvExBench: Benchmarking autonomous exploitation of adversarial example defenses
Carlini, Nicholas
Rando, Javier
Debenedetti, Edoardo
Nasr, Milad
Tramèr, Florian
Cryptography and Security
Artificial Intelligence
Machine Learning
We introduce AutoAdvExBench, a benchmark to evaluate if large language models (LLMs) can autonomously exploit defenses to adversarial examples. Unlike existing security benchmarks that often serve as proxies for real-world tasks, bench directly measures LLMs' success on tasks regularly performed by machine learning security experts. This approach offers a significant advantage: if a LLM could solve the challenges presented in bench, it would immediately present practical utility for adversarial machine learning researchers. We then design a strong agent that is capable of breaking 75% of CTF-like ("homework exercise") adversarial example defenses. However, we show that this agent is only able to succeed on 13% of the real-world defenses in our benchmark, indicating the large gap between difficulty in attacking "real" code, and CTF-like code. In contrast, a stronger LLM that can attack 21% of real defenses only succeeds on 54% of CTF-like defenses. We make this benchmark available at https://github.com/ethz-spylab/AutoAdvExBench.
title AutoAdvExBench: Benchmarking autonomous exploitation of adversarial example defenses
topic Cryptography and Security
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2503.01811