Do Not Merge My Model! Safeguarding Open-Source LLMs Against Unauthorized Model Merging

2025-11-13

AI Summary

Open-source large language models (LLMs) face unauthorized model merging—a form of intellectual property theft—where adversaries combine stolen models to reconstruct or enhance proprietary capabilities. Method: We propose MergeBarrier, the first defense framework that is proactive, open-source compatible, and incurs zero performance degradation. Leveraging linear mode connectivity analysis, MergeBarrier introduces a plug-and-play protection module that injects targeted perturbations into the weight space, thereby disrupting low-loss linear interpolation paths between models and preventing illicit merging. Contribution/Results: Evaluated on mainstream open-source LLMs (e.g., Llama and Qwen series), MergeBarrier effectively thwarts diverse merging attacks while preserving full fine-tuning and inference performance (accuracy drop <0.3%). It seamlessly integrates with standard deployment pipelines (e.g., Hugging Face Transformers). Crucially, this work pioneers the formalization of connectivity disruption as a deployable, proactive defense mechanism—establishing a novel paradigm for copyright protection of open-source LLMs.

Technology Category

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

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Model merging has emerged as an efficient technique for expanding large language models (LLMs) by integrating specialized expert models. However, it also introduces a new threat: model merging stealing, where free-riders exploit models through unauthorized model merging. Unfortunately, existing defense mechanisms fail to provide effective protection. Specifically, we identify three critical protection properties that existing methods fail to simultaneously satisfy: (1) proactively preventing unauthorized merging; (2) ensuring compatibility with general open-source settings; (3) achieving high security with negligible performance loss. To address the above issues, we propose MergeBarrier, a plug-and-play defense that proactively prevents unauthorized merging. The core design of MergeBarrier is to disrupt the Linear Mode Connectivity (LMC) between the protected model and its homologous counterparts, thereby eliminating the low-loss path required for effective model merging. Extensive experiments show that MergeBarrier effectively prevents model merging stealing with negligible accuracy loss.

Problem

• Preventing unauthorized merging of open-source large language models
• Safeguarding against model merging stealing by free-riders
• Disrupting Linear Mode Connectivity to eliminate merging paths

Innovation

• Prevents unauthorized model merging via plug-and-play defense
• Disrupts Linear Mode Connectivity between protected models
• Ensures high security with negligible performance loss