核心概念
Token entropy is crucial in watermark detection, leading to the development of an Entropy-based Watermark Detection (EWD) algorithm.
摘要
In this work, an Entropy-based Text Watermarking Detection Method is proposed to address the challenges of watermark detection in low-entropy scenarios. The method assigns weights to tokens based on their entropy, improving detection accuracy. The theoretical analysis compares EWD with previous methods and validates its performance in code generation tasks. Experiments show superior detection accuracy in low-entropy scenarios while maintaining efficiency.
-
Abstract
- Proposed Entropy-based Watermark Detection (EWD) for text generated by large language models.
- Weight adjustment based on token entropy improves detection performance.
- Training-free and automated process applicable to texts with different entropy distributions.
-
Introduction
- Advancements in large language models pose risks of misuse, necessitating effective watermarking algorithms.
- Text watermarking embeds hidden features for subsequent detection, mitigating misuse risks.
-
Watermarked High Entropy Text
- Probability of drawing top card from a deck explained.
-
Watermarked Low Entropy Code
- Code snippet provided with explanation of low-entropy scenario challenges.
-
Entropy Tag
- Different levels of entropy explained: Low, Mid, High.
-
Data Extraction
- Z-scores provided for different scenarios: 2.54 and 9.29
-
Quotations
- "The influence of token entropy should be fully considered in the watermark detection process."
-
Methodology
- Proposed EWD assigns importance weight to tokens proportional to their entropy for accurate watermark level reflection.
-
Theoretical Analysis
- Type-I and Type-II error analysis conducted comparing EWD with KGW and SWEET methods.
-
Experiments
- Evaluation conducted on code generation tasks showing improved detection accuracy with EWD compared to baselines.
-
Conclusion
- EWD offers a promising solution for watermark detection in low-entropy scenarios with efficient computational cost.
統計資料
Z-score: 2.54
Z-score: 9.29
引述
"The influence of token entropy should be fully considered in the watermark detection process."