New Watermarking Technique Detects Tampering in LLM-Generated Text
A novel watermarking technique for large language models (LLMs) has been unveiled, aimed at tracing origin and identifying tampering simultaneously. Detailed in a paper on arXiv (2608.12713), this method tackles the issue of piggyback spoofing, where current watermarks permit adversaries to modify essential content while still attributing it. The innovative watermark integrates a robust signal alongside a fragile one into each token generated, utilizing distinct keys and varying seeding windows over normalized text. This approach ensures one signal withstands edits while the other reacts to visible changes. The method incorporates multiple rounds of unbiased tournament reweighting to maintain the expected generation distribution, with a periodic allocation pattern managing the balance between signals. Upon detection, scores create a two-dimensional space that allows for three outcomes: Intact, Tampered, and No-Watermark. This advancement is crucial for safeguarding the integrity of AI-generated materials, especially where authenticity and provenance are vital.
Key facts
- The watermark provides both provenance and tamper evidence.
- It co-embeds a robust signal and a fragile signal into each generated token.
- The signals use independent keys and different seeding windows over normalized text.
- Multiple rounds of unbiased tournament reweighting preserve the expected generation distribution.
- A periodic round-allocation pattern controls the trade-off between the two signals.
- Detection scores form a two-dimensional space supporting three decisions: Intact, Tampered, and No-Watermark.
- The method addresses piggyback spoofing, a vulnerability in existing LLM watermarks.
- The paper is available on arXiv with identifier 2608.12713.
Entities
Institutions
- arXiv