RLSpoofer: Distributional Analysis Reveals Fundamental Limits of LLM Watermark Spoofing Resilience

arXiv·low signalXBlueskyLinkedInCopy link

Huang et al. study watermark resilience against spoofing from a distributional perspective, establishing a local capacity bound and introducing RLSpoofer, a lightweight evaluator that quantifies spoofing vulnerability without requiring white-box access to watermark internals. As AI-generated text detection becomes a regulatory requirement, understanding the theoretical limits of watermark robustness is critical — this paper shows that some watermarking schemes are fundamentally more spoofable than others.