Securing LLM Ownership in the Era of Model Merging
Novel fingerprinting technique resists unauthorized model merging
MergePrint introduces a robust fingerprinting system that verifies LLM ownership even after models have been merged with others, addressing a critical intellectual property vulnerability.
- Creates merge-resistant fingerprints that survive when proprietary models are combined with others
- Demonstrates high verification accuracy while maintaining model performance
- Provides black-box verification without requiring access to model parameters
- Outperforms existing fingerprinting methods against merging attacks
As model merging becomes a popular technique to combine capabilities of multiple LLMs, this research addresses a critical security gap for organizations investing in proprietary AI development, offering practical protection against a novel form of IP theft.