Paper Image

Safeguarding AI text generation

Published on:

20 February 2024

Primary Category:

Computation and Language

Paper Authors:

Jiayi Fu,

Xuandong Zhao,

Ruihan Yang,

Yuansen Zhang,

Jiangjie Chen,

Yanghua Xiao

Bullets

Key Details

Proposes GumbelSoft watermark for machine text authentication

GumbelSoft balances detectability and diversity via softmax

Surpasses alternatives in detectability and robustness

Maintains low perplexity like existing methods

Identifies causes of identical outputs in GumbelMax watermarks

AI generated summary

Safeguarding AI text generation

This paper proposes a new GumbelMax-based watermark to authenticate machine-generated texts while enhancing diversity. The GumbelSoft watermark balances detectability and variability through softmax sampling. Comparative analysis shows superior performance.

Answers from this paper

Comments

No comments yet, be the first to start the conversation...

Sign up to comment on this paper

Sign Up