FDLLM: A Dedicated Detector for Black-Box LLMs Fingerprinting

The proliferation of black-box Large Language Models (LLMs) makes source attribution essential for accountability and security. Yet, progress is limited by the lack of a large multilingual benchmark and by fragile or computationally intensive methods. We introduce FD-Dataset, a bilingual benchmark of 90,000 samples from 20 major LLMs, and FDLLM, a LoRA-adapted detector that extracts persistent decoding fingerprints from a foundation model. LoRA induces intra-model clustering and inter-model separation in representation space, explaining its effectiveness for fingerprinting. On FD-Dataset, FDLLM surpasses the strongest baseline by 22.1% Macro F1, generalizes to newly released models with 95% accuracy, and remains robust to polishing, translation, and synonym substitution, reducing average attack success rate from 49.2% (LM-D) to 23.9%.

Paper

Similar papers

© 2026 NYSGPT2525 LLC