On Google's SynthID-Text LLM Watermarking System: Theoretical Analysis and Empirical Validation

Researchers have conducted the first theoretical analysis of Google's SynthID-Text watermarking system, revealing vulnerabilities in its detection methods and proposing attacks that can break the system. The study identifies weaknesses in the mean score detection approach and demonstrates that the Bayesian score offers better robustness, while establishing optimal parameters for watermark detection.

• →Google's SynthID-Text is the first production-ready watermarking system for large language models using a novel Tournament-based sampling method.
• →Researchers discovered that the mean score detection method is vulnerable to layer inflation attacks that can break the watermarking system.
• →The Bayesian score approach provides improved watermark robustness compared to the mean score method.
• →The optimal Bernoulli distribution parameter for watermark detection is 0.5, as proven through theoretical analysis.
• →The research opens new pathways for both watermark removal strategies and designing more robust watermarking techniques.