arXiv · cs.AI · Romina Omidi, Nima Naderloui, Yuan Hong, Binghui Wang · 9 小时前模型评分61
研究提出针对大语言模型统计生成水印的统一攻击框架
A Unified Attack Framework against Statistical Generative Watermarks for Large Language Models
导读
研究提出一个统一攻击框架，针对 KGW 及其变体、GumbelSoft、DipMark、ITS、SynthID 等主流大语言模型统计水印。作者指出这些方案的检测分数与生成时的 token 概率偏移存在关联，称为分数比对齐，攻击者只需黑盒查询、无需密钥或模型权重即可生成规避检测且流畅自然的文本。在七种水印方案、三个大语言模型及相应任务上，检测率在所有设置下被降至最高 0.05，同时保持文本质量。
分类置信度
中 · 待人工复核
- 只有单一媒体来源，需要人工复核。
来源级别：媒体与个人
事件类型：研究发布
证据摘录
- “In this work, we identify a shared weakness across a broad class of popular LLM watermarking methods, including KGW (and its variants), GumbelSoft, DipMark, ITS, and SynthID.”
- “allows an attacker with only black-box query access to the model--and no knowledge of secret keys or model weights--to produce text that evades detection while remaining fluent and natural”
- “We evaluate our attack on seven popular generative LLM watermarking schemes, three LLMs, and tasks, which reduce detection rates to at most 0.05 across all settings while preserving text quality”
来源：arXiv · cs.AI · arxiv.org