FLIPS: Instance-Fingerprinting for LLMs via Pseudo-random Sequences
Researchers have developed FLIPS, a new method for instance-level fingerprinting of Large Language Models (LLMs). This technique exploits biases in generated pseudo-random sequences to identify specific configurations of an LLM, achieving high accuracy in distinguishing between different instances. FLIPS is designed to aid AI regulation by assessing deployed behaviors rather than just model provenance, demonstrating its necessity and feasibility for compliance. AI
IMPACT Enables more granular AI regulation by distinguishing specific deployed LLM configurations.