Anchor-ECC
Localizing suspicious regions in edited watermarked LLM outputs.
A document can retain a strong global watermark even after a small but consequential edit. This project extends watermark verification from document-level attribution to local integrity checking.
Given only the observed text and the watermark key, the method parses structural watermark blocks, tests their coding-theoretic consistency, and flags suspicious regions for downstream review.
- Localizes candidate insertions, deletions, and substitutions
- Uses boundary anchors and feasible-codeword checks
- Submitted to Transactions on Machine Learning Research (TMLR)
- Read the paper on arXiv
- View project website
Watermark generation: tokens carry code symbols and boundary anchors that define locally verifiable blocks.
Local verification: the edited text is parsed into structural blocks, then inconsistent blocks are flagged for review.