Combating Data Laundering in LLM Training

The paper introduces a novel approach to detect unauthorized training data in large language models (LLMs) that are subject to data laundering, where proprietary data is transformed to obscure its origin. The method, termed Synthesis Data Reversion (SDR), utilizes an auxiliary LLM to create queries that mimic the original data, enabling effective detection even when the laundering transformation is undisclosed. Evaluated against models like Pythia, Llama2, and Falcon on the MIMIR benchmark, SDR demonstrates a reliable mechanism for restoring detection signals, thereby providing practitioners with a robust tool for auditing LLM training data integrity.