Sentient Foundation launches open benchmark for LLM fingerprinting robustness and utility

Sentient Foundation has introduced the Fingerprinting Leaderboard, an open benchmark designed to compare methods for embedding identifying signatures into large language models. These signatures, such as secret query-response pairs, allow for origin verification even after a model has been copied or modified.

The benchmark focuses on three critical dimensions:

• Robustness: Measuring how well fingerprints withstand removal attempts or evasion techniques.
• Model utility: Assessing the degree to which the model's original capability is preserved after fingerprinting.
• Verification: Evaluating the accuracy of provenance checks.

The initiative aims to empirically test the tradeoffs involved in building durable model provenance, specifically examining survival rates through fine-tuning, distillation, and deliberate attacks rather than assuming durability.