Making Open-Source Text LLM Watermarks Durable Against Merging
Quick Answer
This paper shows that This research introduces Merge-Adversarial Training to create durable watermarks for open-source LLMs, which resist removal during model merging.
Quick Take
The approach significantly enhances watermark detection performance, achieving up to +51 percentage points in TPR@1%FPR compared to existing methods, while maintaining model capabilities. This advancement is crucial for preserving the integrity of generated text in collaborative AI environments.
Key Points
- Merge-Adversarial Training distills watermarks into model weights effectively.
- Achieves up to +51 percentage points TPR@1%FPR improvement over baselines.
- Evaluates watermark durability against realistic model merging scenarios.
- Preserves downstream capabilities while enhancing watermark robustness.
- Demonstrates adversarial training's reliability for watermark durability.
DeepSignal Analysis
What happened
This research presents Merge-Adversarial Training, a method designed to create durable watermarks for open-source language models (OSMs). The proposed technique enhances watermark detection performance significantly, achieving up to a 51 percentage point increase in true positive rate at a 1% false positive rate compared to existing methods, while maintaining the model's capabilities.
Key evidence
- The study demonstrates that traditional watermarks in OSMs can be removed through post-training modifications, particularly during model merging.
- Merge-Adversarial Training is shown to outperform baseline methods, achieving an average increase of 25 percentage points in TPR@1%FPR.
- The research evaluates OSM watermarks against realistic merging scenarios, which include combining expert capabilities and preventing catastrophic forgetting.
Why it matters
The ability to maintain watermarks in OSMs is crucial for ensuring the integrity of generated text, especially in collaborative AI environments. As OSMs become more prevalent, the durability of these watermarks can help prevent misuse and ensure accountability in AI-generated content. This advancement could lead to more reliable applications of AI in sensitive areas where text authenticity is paramount.
Paper Resources
Source Excerpt
Open-source (OSMs)arereaching near state-of-the-art performance, prompting prior works to trace the text they generate by embedding text watermarking algorithms directly into their weights. Yet, OSMs are subject to post-training modifications, which has been shown to remove the watermark. Model merging in particular, a prominent method used for combining expert knowledge and preventing catastrophic forgetting, strongly removes such OSM watermarks. A key question is how to enable OSM waterma
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from arXiv cs.CL
See more →TriAgent: Divergence-Aware Committees for Cost-Efficient Financial Sentiment Analysis
TriAgent introduces a cost-efficient multi-agent system for financial sentiment analysis, combining VADER, FinBERT, and Qwen2.5. It achieves an F1 score of ~0.87 with significant savings of $9.3M/year at a 10M-user scale compared to GPT-4o-mini, while also detecting hallucinations with an AUC of 0.90.