Reproduce “Fooling the Watchers: Breaking AIGC Detectors via Semantic Prompt 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

Run Hao, Peng Ying

Researchers from Aarhus University and China University of Mining and Technology develop an automated adversarial prompt generation framework that systematically creates text prompts to make AI-generated images bypass AIGC detectors, achieving first place in Tencent's Zhuque AIGC detection competition and demonstrating consistent evasion success against both commercial and open-source detectors through semantic attacks based on texture injection ("clear text") and lighting manipulation ("dazzle") rather than pixel-level perturbations, using a grammar tree structure with UCT-RAND optimization to efficiently explore the vast prompt space and generate diverse adversarial examples that maintain photorealistic quality.

Key insights

  • Semantic prompt attacks represent a fundamentally different vulnerability than pixel-level adversarial perturbations, operating through the T2I model's natural control mechanisms
  • Advanced T2I models with higher parameter counts are more easily detected by PatchCraft, while weaker models naturally evade detection, suggesting training data limitations in current detectors
  • Grammar tree structure enables systematic exploration of prompt semantics while UCT-RAND optimization efficiently discovers successful evasion strategies
  • Current AIGC detectors rely heavily on low-level artifacts and are inadequately prepared for semantic-level manipulations

Summary by alphaXiv.

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