Tracing Provenance and Detecting Tampering with Complementary LLM Watermarks

Abstract

Watermarking LLM-generated text is an important task for tracing its provenance. Existing LLM watermarks preserve provenance under editing, but this same robustness allows an adversary to alter critical content while retaining attribution, a vulnerability known as piggyback spoofing. We introduce an innovative watermark that jointly provides provenance and tamper evidence. It co-embeds a robust signal and a fragile signal into each generated token. The signals share the same mechanism but use in

Categories

@misc{feng2026tracing,
  title = {{Tracing Provenance and Detecting Tampering with Complementary LLM Watermarks}},
  author = {Xiao-Yang Feng and Yanjun Zhang and He Zhang and L. Zhang and Shirui Pan},
  year = {2026},
  month = aug,
  eprint = {2608.12713},
  archivePrefix = {arXiv},
  url = {https://www.semanticscholar.org/paper/ab116a93501cd418ff258dda5081b8daf086b858}
}