Learning to Detect Language Model Training Data via Active Reconstruction

Large Language Models (LLMs) utilize large amounts of data for their training, some of which may come from copyrighted sources. Membership Inference Attacks (MIA) aim to detect those documents and whether they have been included in the training corpora of the LLMs. The black-box MIAs require a significant amount of data manipulation; therefore, their comparison is often challenging. We study state-of-the-art (SOTA) MIAs under the black-box assumptions and compare them to each other using a unified set of datasets to determine if any of them can reliably detect membership under SOTA LLMs. In addition, a new method, called the Familiarity Ranking, was developed to showcase a possible approach to black-box MIAs, thereby giving LLMs more freedom in their expression to understand their reasoning better. The results indicate that none of the methods are capable of reliably detecting membership in LLMs, as shown by an AUC-ROC of approximately 0.5 for all methods across several LLMs. The higher TPR and FPR for more advanced LLMs indicate higher reasoning and generalizing capabilities, showcasing the difficulty of detecting membership in LLMs using black-box MIAs.

Post-hoc unauthorized-training data detection for large language models (LLMs) typically assumes a query-with-originals regime: rights holders query a target LLM with raw proprietary data and assess whether the model assigns them stronger memorization-based detection signals, e.g., higher confidence or lower loss, than held-out non-training reference texts. We show that this regime becomes brittle under data laundering, where the target LLM is trained on semantics-preserving but stylistically or structurally transformed surrogates of proprietary data to obfuscate provenance. Since training-time exposure occurs in the laundered form, memorization signals may no longer appear on the originals, collapsing the candidate-reference signal separation that standard detectors rely on. We counter this threat by studying laundering-aware detection with raw proprietary data, a held-out reference corpus, and query access to the target LLM, while the laundering transformation is undisclosed. Since exact recovery of the laundered corpus is infeasible, we infer a detection-useful synthesis process via an auxiliary LLM that maps originals into training-like queries. To make this search tractable, we introduce Synthesis Data Reversion (SDR), which constrains the unbounded space of natural-language transformations through a goal-details abstraction: a high-level transformation goal, e.g., "lyrical rewriting", and fine-grained details, e.g., "with vivid imagery". SDR identifies the most likely goal and iteratively refines details so synthesized queries elicit stronger target-model detection signals. Evaluated on the MIMIR benchmark against diverse laundering practices and target LLM families (Pythia, Llama2, and Falcon), SDR consistently restores detection signals, offering a practical auditing layer against data laundering.

Membership inference on large language models (LLMs) aims to determine whether a given text sample was included in an LLM's training data, without access to its training corpus. Despite recent progress, existing benchmarks suffer from three limitations: limited coverage of training stages, insufficient distributional alignment between members and non-members, and lack of rigorous filtering of non-members against the training corpus. To address these limitations, we propose OLMo-Detect, a multi-stage, confounder-controlled benchmark built upon the fully open OLMo 2 pipeline. OLMo-Detect spans pre-training, mid-training, and post-training, explicitly aligns members and non-members on three key axes, and rigorously filters non-members via infini-gram. To assess robustness to distribution shifts, we further introduce OLMo-Detect (Shifted), a variant where members are misaligned with non-members. We evaluate 15 unsupervised and 3 supervised membership inference attacks (MIAs) across the OLMo 2 family, finding that: (i) overall performance is limited: the best unsupervised and supervised MIAs both reach an AUC of only 0.68, and supervised MIAs degrade under cross-domain evaluation; (ii) MIA performance peaks at mid-training and is lower at pre-training and post-training, a pattern driven by data type rather than a stage effect: curated math data is far more detectable than other types; (iii) overall scores improve from 1B to 13B but plateau at 32B; and (iv) no unsupervised MIA is robust to distribution shifts, with AUCs shifting by up to 0.42. Finally, we find that our findings on OLMo 2 generalize to OLMo 3 and non-OLMo models.