SWAN: Semantic Watermarking via Abstract Meaning Representation
A new framework called SWAN (Semantic Watermarking with Abstract Meaning Representation) embeds watermark signatures into the semantic structure of sentences using AMR. Unlike token-level methods, SWAN encodes signatures at the semantic level, making them robust to paraphrasing. The method is training-free: injection uses an LLM guided by AMR templates, detection uses an off-the-shelf AMR parser and a z-test. Evaluated on RealNews, SWAN matches state-of-the-art detection performance.
Key facts
- SWAN stands for Semantic Watermarking with Abstract Meaning Representation.
- Watermark signatures are embedded into the semantic structure of a sentence using AMR.
- Existing methods encode signatures by adjusting token selection during text generation.
- SWAN encodes the signature directly in the sentence's semantic representation.
- Any paraphrase that preserves meaning automatically preserves the signature.
- SWAN is training-free: injection uses prompting an LLM with selected AMR templates.
- Detection uses an off-the-shelf AMR parser followed by a one-proportion z-test.
- Empirical evaluation on the RealNews benchmark shows SWAN matches state-of-the-art detection performance.
Entities
—