SemTrace: Source-Grounded Semantic Signatures for Tracing LLM Exposure to Protected Documents
SemTrace proposes tracing protected-document exposure through semantic facts embedded in an assigned manuscript copy.
The paper describes a watermark that does not depend on token patterns or access to the model generating the review. It builds a document-specific binary signature from factual propositions supported by the manuscript, then has a protected PDF carry instructions that steer selected facts into fixed review slots. A frozen natural language inference model decodes those semantic signals and checks them against the copy’s codeword. ArXiv · AI/CL/LG's note
The paper describes a watermark that does not depend on token patterns or access to the model generating the review. It builds a document-specific binary signature from factual propositions supported by the manuscript, then has a protected PDF carry instructions that steer selected facts into fixed review slots. A frozen natural language inference model decodes those semantic signals and checks them against the copy’s codeword. ArXiv · AI/CL/LG's note
score 5