MergePrint: Merge-Resistant Fingerprints for Robust Black-box Ownership Verification of Large Language Models

TL;DR

• LLM fingerprints survive model merging: MergePrint optimizes fingerprints against a pseudo-merged model.
• Ownership can be verified through queries alone: Checks target outputs without accessing model weights or intermediate features.
• Fingerprinting limits damage to model utility: Pre-optimizes fingerprint inputs to reduce performance degradation.

Futa Waseda

Project Assistant Professor