Combating Data Laundering in LLM Training: Detecting Stylistically Transformed Proprietary Data

arXiv·medium signal

Researchers tackle data laundering — the practice of transforming the stylistic form of proprietary data while preserving its semantic content to evade membership inference detection. Standard detection methods become fragile when training data has been paraphrased or reformatted. The paper presents techniques to detect unauthorized data use even after stylistic transformation, relevant to data rights owners trying to enforce licensing terms against LLM providers.