Reproduce “What the Eyes See, the LLMs Miss: Exploiting Human Perception for Adversarial Text Attacks” 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

Qin Yang, Lu Malloy, Joshua Lee, Xiaohan Chang, Meisam Mohammady, Doowon Kim, Yuan Hong

A new class of adversarial attacks, Human-Perceptible Adversarial Attacks (HPAA), was developed to leverage the difference between how humans visually interpret text and how automated moderation systems process tokenized input. This approach enabled up to 100% evasion of harmful content by 13 commercial and open-source moderation systems, simultaneously ensuring over 86% human recognition of the embedded toxicity.

Key insights

  • A fundamental "blind spot" exists in LLM-based moderation due to its reliance on tokenized text, which ignores visual cues critical for human interpretation.
  • Specific typographic configurations can effectively embed harmful content, making it highly recognizable to human readers despite appearing benign to machines.
  • Integrating user studies into the attack design process is essential for systematically exploring and exploiting the human-machine perceptual mismatch in content moderation.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.