Reproduce “RAFT: Realistic Attacks to Fool Text Detectors” 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
James Wang, Ran Li, Junfeng Yang, Chengzhi Mao
Researchers from Columbia and Rutgers developed RAFT, an adversarial attack that can subvert state-of-the-art LLM detectors by up to 99% AUROC reduction, generating text that remains grammatically correct, semantically consistent, and human-indistinguishable. This black-box framework preserves linguistic quality while demonstrating critical vulnerabilities in existing detection mechanisms.
Key insights
- Existing LLM detection methods are highly vulnerable to realistic, human-indistinguishable adversarial attacks, with AUROC reductions reaching up to 99%.
- Perturbations maintaining Part-of-Speech consistency and semantic fluency are crucial for crafting high-quality, practically undetectable adversarial examples.
- Adversarial examples generated by RAFT can be effectively used for adversarial training to improve the robustness of LLM detectors, enabling them to better resist future attacks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.