Invisible Entropy: Towards Safe and Efficient Low-Entropy LLM Watermarking

2025-05-20

AI Summary

Existing LLM watermarking methods suffer from degraded text naturalness and poor detection robustness in low-entropy generation scenarios; their reliance on original-model entropy computation incurs high computational overhead, detection latency, and model leakage risks. Method: We propose a model-free, lightweight watermarking paradigm: (i) a novel entropy prediction mechanism leveraging a lightweight feature extractor and entropy annotator—requiring no access to the original LLM; and (ii) a theory-driven adaptive threshold navigator that dynamically modulates green/red token distributions in the logit space to jointly preserve textual fluency and enhance detection robustness. Contribution/Results: Our method reduces parameter count by 99%, achieves state-of-the-art detection performance on HumanEval and MBPP, and fully eliminates model dependency and associated computational costs.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Computer Vision: Large Vision Models

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Logit-based LLM watermarking traces and verifies AI-generated content by maintaining green and red token lists and increasing the likelihood of green tokens during generation. However, it fails in low-entropy scenarios, where predictable outputs make green token selection difficult without disrupting natural text flow. Existing approaches address this by assuming access to the original LLM to calculate entropy and selectively watermark high-entropy tokens. However, these methods face two major challenges: (1) high computational costs and detection delays due to reliance on the original LLM, and (2) potential risks of model leakage. To address these limitations, we propose Invisible Entropy (IE), a watermarking paradigm designed to enhance both safety and efficiency. Instead of relying on the original LLM, IE introduces a lightweight feature extractor and an entropy tagger to predict whether the entropy of the next token is high or low. Furthermore, based on theoretical analysis, we develop a threshold navigator that adaptively sets entropy thresholds. It identifies a threshold where the watermark ratio decreases as the green token count increases, enhancing the naturalness of the watermarked text and improving detection robustness. Experiments on HumanEval and MBPP datasets demonstrate that IE reduces parameter size by 99% while achieving performance on par with state-of-the-art methods. Our work introduces a safe and efficient paradigm for low-entropy watermarking. https://github.com/Carol-gutianle/IE https://huggingface.co/datasets/Carol0110/IE-Tagger

Problem

• Overcoming low-entropy challenges in LLM watermarking
• Reducing computational costs and model leakage risks
• Enhancing watermarking safety and efficiency adaptively

Innovation

• Lightweight feature extractor predicts token entropy
• Adaptive entropy threshold enhances text naturalness
• 99% parameter reduction with state-of-the-art performance