Three bricks to consolidate watermarks for large language models - Inria - Institut national de recherche en sciences et technologies du numérique
Conference Papers Year : 2023

Three bricks to consolidate watermarks for large language models

Abstract

Discerning between generated and natural texts is increasingly challenging. In this context, watermarking emerges as a promising technique for ascribing text to a specific generative model. It alters the sampling generation process to leave an invis- ible trace in the output, facilitating later detection. This research consolidates watermarks for large language models based on three theoretical and empirical considerations. First, we introduce new statistical tests that offer robust theoretical guarantees which remain valid even at low false-positive rates (less than 10-6). Second, we compare the effectiveness of watermarks using classical benchmarks in the field of natural language processing, gaining insights into their real-world applicability. Third, we develop advanced detection schemes for scenarios where access to the LLM is available, as well as multi-bit watermarking.
Fichier principal
Vignette du fichier
fernandez_wifs2023.pdf (773.99 Ko) Télécharger le fichier
Origin Files produced by the author(s)

Dates and versions

hal-04361015 , version 1 (22-12-2023)

Licence

Identifiers

  • HAL Id : hal-04361015 , version 1

Cite

Pierre Fernandez, Antoine Chaffin, Karim Tit, Vivien Chappelier, Teddy Furon. Three bricks to consolidate watermarks for large language models. WIFS 2023 - IEEE International Workshop on Information Forensics and Security, Dec 2023, Nuremberg, Germany. pp.1-9. ⟨hal-04361015⟩
96 View
64 Download

Share

More