Researchers have developed a method to analyze the robustness of watermarks in open-weight large language models against modifications. The study derives theoretical bounds for watermark evidence under various editing and fine-tuning attacks, providing a better understanding of how these watermarks perform when the model weights are altered. Experiments using OpenStamp on LLaMA-2-7B and Mistral-7B-v0.3 demonstrated that while quantization and pruning had minimal impact, fine-tuning significantly reduced watermark detection rates. AI
IMPACT Provides a theoretical framework and empirical evidence for understanding the resilience of LLM watermarks to model modifications, informing future watermark design and security.
RANK_REASON The cluster contains a research paper detailing theoretical and empirical analysis of LLM watermark robustness. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →