Auditing Training Data in Domain-adapted LLMs: LoRA-MINT

In three linesLoRA-MINT is an auditing method to detect whether individual data samples were used in training LoRA-adapted LLMs. Tested on 4 models and 3 benchmarks, it achieves precision of 0.77–0.92 for identifying data exposure, outperforming existing baselines.

Summary generated by Claude — human-verified