Reproduce “FlipJev: Fast Black-Box Red Teaming Finds Universal Adversarial Suffixes for Typed Decision Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xiaogeng Liu, Xinyan Wang, Chaowei Xiao
FlipJev searches Jev’s returned answer-option probabilities to find short suffixes appended to untrusted content, without using an attacker language model. On nine classification datasets, one suffix per dataset flipped 114 of 264 held-out documents (43%); a suffix found on full arXiv papers flipped 90–100% of the first 20 held-out documents across five other datasets. These findings concern Jev under the study’s label-preserving threat model.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.