Keyed homoglyph substitution for LLM text provenance
Original title: KeHoSu: Keyed Homoglyph Substitution for Authenticated Text Provenance over Copy-Paste Channels
KeHoSu encodes attribution in homoglyph variants over copy-paste channels, surviving reordering and partial quotes for accurate platform and account attribution.
LLM-generated text redistributed via copy-paste discards metadata. KeHoSu turns confusable characters into keyed attribution channel encoding platform, user and truncated tag bits. Method invariant to sentence reordering up to boundary layer. Survives copy-paste across 18 platforms, attributing majority of excerpts correctly at 167-178 letters.
- One bit encoded per carrier letter
- Invariant to reordering and partial quotes
- Capacity characterized for removability and detectability
- Achieves majority attribution at ~170 letters
- Publicly evaluable over common platforms
From the original article
1 min read at arxiv.orgNowadays, large language models produce text at scale, and much of it is redistributed across platforms. A moderator who finds a suspicious comment, post or article may want to know which platform generated it and for which account. In this paper we study attribution over the channel through which such text actually travels, copy-paste, which discards document metadata but keeps Unicode code points.
By Andrei Alexandru Drăghici · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.