Topic-based watermarking to identify AI text

Paper Title:

Topic-based Watermarks for LLM-Generated Text

Published on:

2 April 2024

Primary Category:

Cryptography and Security

Paper Authors:

Alexander Nemecek,

Yuzhou Jiang,

Erman Ayday

Bullets

Key Details

•

Proposes topic-based watermarking technique for large language models

•

Embeds signatures in AI text based on extracted topic(s)

•

Overcomes limitations in robustness and practicality of prior schemes

•

Enables feasible watermark detection algorithms at scale

•

Allows modeling attacker's benefit vs. loss tradeoff

Explore the topics in this paper

attack modeling

language models

natural language processing

text generation

watermarking

AI generated summary

Topic-based watermarking to identify AI text

This paper proposes a new watermarking technique to identify text generated by large language models versus humans. It embeds detectable signatures based on the text's topics, overcoming limitations in previous watermarking schemes that lacked robustness against attacks or practicality at scale. The proposed technique selects inclusion/exclusion token lists according to extracted topics, enabling feasible detection algorithms. It provides modeling of potential attacks' benefits versus losses.