Code-Mixing Fingerprints Improve LLM Ownership Verification Under Black-Box Filtering

Code-mixing fingerprints (CF) select triggers at the lowest perplexity under a high-complexity constraint, avoiding both accidental activation of natural-language triggers and the statistical detectability of garbled ones. The end-to-end injection framework is designed to survive downstream model modifications that degrade standard fingerprint triggers.