Detecting Non-Membership in LLM Training Data via Rank Correlations
This addresses the need for copyright enforcement, compliance auditing, and user trust in LLMs by providing a method to verify dataset exclusion, though it is incremental as it builds on prior membership inference work.
The paper tackles the problem of verifying that a dataset was not used in training large language models (LLMs), introducing PRISM, a test that detects dataset-level non-membership using rank correlations in model logits, and it reliably rules out membership across all tested datasets without false positives.
As large language models (LLMs) are trained on increasingly vast and opaque text corpora, determining which data contributed to training has become essential for copyright enforcement, compliance auditing, and user trust. While prior work focuses on detecting whether a dataset was used in training (membership inference), the complementary problem -- verifying that a dataset was not used -- has received little attention. We address this gap by introducing PRISM, a test that detects dataset-level non-membership using only grey-box access to model logits. Our key insight is that two models that have not seen a dataset exhibit higher rank correlation in their normalized token log probabilities than when one model has been trained on that data. Using this observation, we construct a correlation-based test that detects non-membership. Empirically, PRISM reliably rules out membership in training data across all datasets tested while avoiding false positives, thus offering a framework for verifying that specific datasets were excluded from LLM training.