Watermarking Makes Language Models Radioactive — Poster
Poster 'Watermarking Makes Language Models Radioactive' by Tom Sander et al. illustrating detection of LLMs trained on watermarked text using greenlist/redlist methods, featuring a 'Radioactive Watermarks' pipeline and graphs of log10(p-value) versus training data proportion.
论文上下文
Paper title: Watermarking Makes Language Models Radioactive Abstract: Poster 'Watermarking Makes Language Models Radioactive' by Tom Sander et al. illustrating detection of LLMs trained on watermarked text using greenlist/redlist methods, featuring a 'Radioactive Watermarks' pipeline and graphs of log10(p-value) versus training data proportion. Paper body (method & results): <!DOCTYPE html> <html lang="en"> <head> <meta content="text/html; charset=utf-8" http-equiv="content-type"/> <title>Watermarking Makes Language Models Radioactive</title> <!--Generated on Thu Feb 22 18:38:03 2024 by LaTeXML (version 0.8.7) http://dlmf.nist.gov/LaTeXML/.--> <meta content="width=device-width, initial-scale=1, shrink-to-fit=no" name="viewport"/> <link href="https://cdn.jsdelivr.net/npm/bootstrap@5.3.0/dist/css/bootstrap.min.css" rel="stylesheet" type="text/css"/> <link href="/static/browse/0.3.4/css/ar5iv_0.7.4.min.css" rel="stylesheet" type="text/css"/> <link href="/static/browse/0.3.4/css/latexml_styles.css" rel="stylesheet" type="text/css"/> <script src="https://cdn.jsdelivr.net/npm/bootstrap@5.3.0/dist/js/bootstrap.bundle.min.js"></script> <script src="https://cdnjs.cloudflare.com/ajax/libs/html2canvas/1.3.3/html2canvas.min.js"></script> <script src="/static/browse/0.3.4/js/addons.js"></script> <script src="/static/browse/0.3.4/js/feedbackOverlay.js"></script> <base href="/html/2402.14904v1/"/></head> <body> <nav class="ltx_page_navbar"> <nav class="ltx_TOC"> <ol class="ltx_toclist"> <li class="ltx_tocentry ltx_tocentry_section"><a class="ltx_ref" href="https://arxiv.org/html/2402.14904v1#S1" title="1 Introduction ‣ Watermarking Makes Language Models Radioactive"><span class="ltx_text ltx_ref_title"><span class="ltx_tag ltx_tag_ref">1 </span>Introduction</span></a></li> <li class="ltx_tocentry ltx_tocentry_section"> <a class="ltx_ref" href="https://arxiv.org/html/2402.14904v1#S2" title="2 Background ‣ Watermarking Makes Language Models Radioactive"><span class="ltx_text ltx_ref_title"><span class="ltx_tag ltx_tag_ref">2 </span>Background</span></a> <ol class="ltx_toclist ltx_toclist_section"> <li class="ltx_tocentry ltx_tocentry_subsection"> <a class="ltx_ref" href="https://arxiv.org/html/2402.14904v1#S2.SS1" title="2.1 Related work ‣ 2 Background ‣ Watermarking Makes Language Models Radioactive"><span class="ltx_text ltx_ref_title"><span class="ltx_tag ltx_tag_ref">2.1 </span>Related work</span></a> <ol class="ltx_toclist ltx_toclist_subsection"> <li class="ltx_tocentry ltx_tocentry_paragraph"><a class="ltx_ref" href="https://arxiv.org/html/2402.14904v1#S2.SS1.SSS0.Px1" title="Membership inference attacks ‣ 2.1 Related work ‣ 2 Background ‣ Watermarking Makes Language Models Radioactive"><span class="ltx_text ltx_ref_title">Membership inference attacks</span></a></li> <li class="ltx_tocentry ltx_tocentry_paragraph"><a class="ltx_ref" href="https://arxiv.org/html/2402.14904v1#S2.SS1.SSS0.Px2" title="Watermarking for LLMs. ‣ 2.1 Related work ‣ 2 Background ‣ Watermarking Makes Language Models Radioactive"><span class="ltx_text ltx_ref_title">Watermarking for LLMs.</span></a></li> </ol> </li> <li class="ltx_tocentry ltx_tocentry_subsection"><a class="ltx_ref" href="https://arxiv.org/html/2402.14904v1#S2.SS2" title="2.2 Technical background for LLM watermarking ‣ 2 Background ‣ Watermarking Makes Language Models Radioactive"><span class="ltx_text ltx_ref_title"><span class="ltx_tag ltx_tag_ref">2.2 </span>Technical background for LLM watermarking</span></a></li> </ol> </li> <li class="ltx_tocentry ltx_tocentry_section"> <a class="ltx_ref" href="https://arxiv.org/html/2402.14904v1#S3" title="3 Problem Formulation ‣ Watermarking Makes Language Models Radioactive"><span class="ltx_text ltx_ref_title"><span class="ltx_tag ltx_tag_ref">3 </span>Problem Formulation</span></a> <ol class="ltx_toclist ltx_toclist_section"> <li class="ltx_tocentry ltx_tocentry_subsection"><a class="ltx_ref" href="https://arxiv.org/html/2402.14904v1#S3.SS1" title="3.1 Notations and Problem Statement ‣ 3 Problem Formulation ‣ Watermarking Makes Language Models Radioactive"><span class="ltx_text ltx_ref_title"><span class="ltx_tag ltx_tag_ref">3.1 </span>Notations and Problem Statement</span></a></li> <li class="ltx_tocentry ltx_tocentry_subsection"><a class="ltx_ref" href="https://arxiv.org/html/2402.14904v1#S3.SS2" title="3.2 Radioactivity ‣ 3 Problem Formulation ‣ Watermarking Makes Language Models Radioactive"><span class="ltx_text ltx_ref_title"><span class="ltx_tag ltx_tag_ref">3.2 </span>Radioactivity</span></a></li> </ol> </li> <li class="ltx_tocentry ltx_tocentry_section"> <a class="ltx_ref" href="https://arxiv.org/html/2402.14904v1#S4" title="4 Radioactivity Detection ‣ Watermarking Makes Language Models Radioactive"><span class="ltx_text ltx_ref_title"><span class="ltx_tag ltx_tag_ref">4 </span>Radioactivity Detection</span></a> <ol class="ltx_toclist ltx_toclist_section"> <li class="ltx_tocentry ltx_tocentry_subsection"> <a class="ltx_ref" href="https://arxiv.org/html/2402.14904v1#S4.SS1" title="4.1 Without watermarking ‣ 4 Radioactivity Detection ‣ Watermarking Makes Language Models Radioactive"><span class="ltx_text ltx_ref_title"><span class="ltx_tag ltx_tag_ref">4.1 </span>Without watermarking</span></a> <ol class="ltx_toclist ltx_toclist_subsection"> <li class="ltx_tocentry ltx_tocentry_paragraph"><a class="ltx_ref" href="https://arxiv.org/html/2402.14904v1#S4.SS1.SSS0.Px1" title="K-S test. ‣ 4.1 Without watermarking ‣ 4 Radioactivity Detection ‣ Watermarking Makes Language Models Radioactive"><span class="ltx_text ltx_ref_title">K-S test.</span></a></li> </ol> </li> <li class="ltx_tocentry ltx_tocentry_subsection"> <a class="ltx_ref" href="https://arxiv.org/html/2402.14904v1#S4.SS2" title="4.2 With watermarking ‣ 4 Radioactivity Detection ‣ Watermarking Makes Language Models Radioactive"><span class="ltx_text ltx_ref_title"><span class="ltx_tag ltx_tag_ref">4.2 </span>With watermarking</span></a> <ol class="ltx_toclist ltx_toclist_subsection"> <li class="ltx_tocentry ltx_tocentry_paragraph"><a class="ltx_ref" href="https://arxiv.org/html/2402.14904v1#S4.SS2.SSS0.Px1" title="Naive approach. ‣ 4.2 With watermarking ‣ 4 Radioactivity Detection ‣ Watermarking Makes Language Models Radioactive"><span class="ltx_text ltx_ref_title">Naive approach.</span></a></li> <li class="ltx_tocentry ltx_tocentry_paragraph"><a class="ltx_ref" href="https://arxiv.org/html/2402.14904v1#S4.SS2.SSS0.Px2" title="Watermark detection in ℬ. ‣ 4.2 With watermarking ‣ 4 Radioactivity Detection ‣ Watermarking Makes Language Models Radioactive"><span class="ltx_text ltx_ref_title">Watermark detection in <math alttext="\mathcal{B}" class="ltx_Math" display="inline"><semantics><mi class="ltx_font_mathcaligraphic">ℬ</mi><annotation-xml encoding="MathML-Content"><ci>ℬ</ci></annotation-xml><annotation encoding="application/x-tex">\mathcal{B}</annotation><annotation encoding="application/x-llamapun">caligraphic_B</annotation></semantics></math>.</span></a></li> <li class="ltx_tocentry ltx_tocentry_paragraph"><a class="ltx_ref" href="https://arxiv.org/html/2402.14904v1#S4.SS2.SSS0.Px3" title="Filter on scored 𝑘-grams. ‣ 4.2 With watermarking ‣ 4 Radioactivity Detection ‣ Watermarking Makes Language Models Radioactive"><span class="ltx_text ltx_ref_title">Filter on scored <math alttext="k" class="ltx_Math" display="inline"><semantics><mi>k</mi><annotation-xml encoding="MathML-Content"><ci>𝑘</ci></annotation-xml><annotation encoding="application/x-tex">k</annotation><annotation encoding="application/x-llamapun">italic_k</annotation></semantics></math>-grams.</span></a></li> <li class="ltx_tocentry ltx_tocentry_paragraph"><a class="ltx_ref" href="https://arxiv.org/html/2402.14904v1#S4.SS2.SSS0.Px4" title="Token scoring and de-duplication. ‣ 4.2 With watermarking ‣ 4 Radioactivity Detection ‣ Watermarking Makes Language Models Radioactive"><span class="ltx_text ltx_ref_title">Token scoring and de-duplication.</span></a></li> </ol> </li> </ol> </li> <li class="ltx_tocentry ltx_tocentry_section"> <a class="ltx_ref" href="https://arxiv.org/html/2402.14904v1#S5" title="5 Radioactivity in Instruction Datasets ‣ Watermarking Makes Language Models Radioactive"><span class="ltx_text ltx_ref_title"><span class="ltx_tag ltx_tag_ref">5 </span>Radioactivity in Instruction Datasets</span></a> <ol class="ltx_toclist ltx_toclist_section"> <li class="ltx_tocentry ltx_tocentry_subsection"><a class="ltx_ref" href="https://arxiv.org/html/2402.14904v1#S5.SS1" title="5.1 Experimental setup ‣ 5 Radioactivity in Instruction Datasets ‣ Watermarking Makes Language Models Radioactive"><span class="ltx_text ltx_ref_title"><span class="ltx_tag ltx_tag_ref">5.1 </span>Experimental setup</span></a></li> <li class="ltx_tocentry ltx_tocentry_subsection"><a class="ltx_ref" href="https://arxiv.org/html/2402.14904v1#S5.SS2" title="5.2 Quality inspection of the instruction tuning ‣ 5 Radioactivity in Instruction Datasets ‣ Watermarking Makes Language Models Radioactive"><span class="ltx_text ltx_ref_title"><span class="ltx_tag ltx_tag_ref">5.2 </span>Quality inspection of the instruction tuning</span></a></li> <li class="ltx_tocentry ltx_tocentry_subsection"><a class="ltx_ref" href="https://arxiv.org/html/2402.14904v1#S5.SS3" title="5.3 A Baseline with membership inference attacks ‣ 5 Radioactivity in Instruction Datasets ‣ Watermarking Makes Language Models Radioactive"><span class="ltx_text ltx_ref_title"><span class="ltx_tag ltx_tag_ref">5.3 </span>A Baseline with membership inference attacks</span></a></li> <li class="ltx_tocentry ltx_tocentry_subsection"><a class="ltx_ref" href="https://arxiv.org/html/2402.14904v1#S5.SS4" title="5.4 With watermarking: open-model setting ‣ 5 Radioactivity in Instruction Datasets ‣ Watermarking Makes Language Models Radioactive"><span class="ltx_text ltx_ref_title"><span class="ltx_tag ltx_tag_ref">5.4 </span>With watermarking: open-model setting</span></a></li> <li class="ltx_tocentry ltx_tocentry_subsection"><a class="ltx_ref" href="https://arxiv.org/html/2402.14904v1#S5.SS5" title="5.5 With watermarking: closed-model setting ‣ 5 Radioactivity in Instruction Datasets ‣ Watermarking Makes Language Models Radioactive"><span class="ltx_text ltx_ref_title"><span class="ltx_tag ltx_tag_ref">5.5 </span>With watermarking: closed-model setting</span></a></li> <li class="ltx_tocentry ltx_tocentry_subsection"><a class="ltx_ref" href="https://arxiv.org/html/2402.14904v1#S5.SS6" title="5.6 Summary & Discussion ‣ 5 Radioactivity in Instruction Datasets ‣ Watermarking Makes Language Models Radioactive"><span class="ltx_text ltx_ref_title"><span class="ltx_tag ltx_tag_ref">5.6 </span>Summary & Discussion</span></a></li> </ol> </li> <li class="ltx_tocentry ltx_tocentry_section"> <a class="ltx_ref" href="https://arxiv.org/html/2402.14904v1#S6" title="6 Investigating Radioactivity ‣ Watermarking Makes Language Models Radioactive"><span class="ltx_text ltx_ref_title"><span class="ltx_tag ltx_tag_ref">6 </span>Investigating Radioactivity</span></a> <ol class="ltx_toclist ltx_toclist_section"> <li class="ltx_tocentry ltx_tocentry_subsection"><a class="ltx_ref" href="https://arxiv.org/html/2402.14904v1#S6.SS1" title="6.1 Fine-tuning ‣ 6 Investigating Radioactivity ‣ Watermarking Makes Language Models Radioactive"><span class="ltx_text ltx
