acceptodds
Under review as a conference paper at ICLR 2027

A False Sense of Provenance: Data Laundering Defeats Training-Data Licensing Audits

Abstract

Training data is now bought and sold under contract. To check these contracts after the fact, researchers have built data-provenance audits, which look for a signal of the licensed data in a trained model. In this paper, we show that a licensee can train on the licensed dataset beyond what its license covers and defeat these audits with a single cheap operation, which we call data laundering. The licensee fine-tunes a single generator on the licensed dataset, regenerates the dataset from that generator, and trains each deployed model on the regenerated dataset. Any signal an audit looks for is either separable from the capability the licensee paid for or inseparable from it, and neither case attributes a laundered model to the licensed dataset. We test the attack against nine published state-of-the-art audits over images and text. Laundered models score at or near the level of models trained without the licensed data. Language models trained on the regenerated dataset keep 82% to 85% of the capability gain from training on the licensed dataset. Laundering also defeats six defenses, including two published audits designed with regeneration in mind.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
data licensing, data provenance, data laundering, membership inference, dataset inference, watermarking, text-to-image models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_i82pOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes data laundering, an attack in which a licensee fine-tunes a generator on a licensed dataset, regenerates a synthetic version of it, and trains deployed models on the regenerated data. Evaluated against nine state-of-the-art audits, laundered models score at or near the level of models trained without the licensed data, while language models still retain 82%–85% of the capability gain from the original dataset.

Critical strengths

The motivation of this paper is sound: it aims to address the problem of data laundering, which is increasingly important in an era of rapidly advancing large models such as LLMs, where the community has reached a consensus that high-quality data matters more and more.

Critical weaknesses

Weakness 1: The experiments are far from sufficient. For images, a licensee could regenerate data through many different operations, such as rotation, style transfer, resizing, or color transformation. I believe that these transformations, which vary in both granularity and modality, pose substantially different challenges and difficulties for data-licensing audits. However, I did not find sufficiently fine-grained experiments that analyze these fundamental factors one by one.

Weakness 2: The situation becomes even less convincing for text. Language can be rewritten at the word level, or augmented at the semantic level without changing the original meaning, both of which can be regarded as important instantiations of regeneration. Therefore, I do not believe that the current text experiments are sufficient to demonstrate the effectiveness of the proposed attack.

Weakness 3: The field of fingerprinting is also advancing rapidly, with a growing body of work on natural fingerprints. I believe that if such natural fingerprints can be well preserved in the regenerated dataset, then the significance of auditing is substantially weakened, since one could simply rely on natural fingerprints as the basis for attribution.

Questions for authors

See Weaknesses

2 comments
AuthorsOpenReview ↗

Response to Reviewer i82p

We thank the reviewer for the insightful comments. We are committed to making all necessary changes and improvements. We performed the three experiments the reviewer suggested and added the results to the revised paper (in blue).

Fine-grained analysis of image transforms (W1)

We agree with the reviewer that the paper tested its membership audit (CLiD scores with CDI's t-test, Sec. 4.1) only on standard data augmentation (crops, flips and color jitter, App. E.4). However, new tests on 17 settings of rotation, resizing, color change and style transfer (Gatys et al., 2016) show that the audit still detects all 102 models trained on transformed images (p < 0.001), including those at the strongest of these settings (30° rotation, resizing to 25%, grayscale, the highest style weight). The average AUC per transform is between 0.89 and 0.99, against 0.42 for the innocent models in Table 3. This is because a transform edits each licensed image in place, so the model trains on all the licensed images. Regeneration instead trains the model only on a dataset D′ sampled from a generator fine-tuned on D (Sec. 3.2).

We added the results for each setting to Sec. 4.3 and App. E.7 (Table 14).

Word-level and semantic text rewrites (W2)

We agree with the reviewer that the paper compared laundering with only one semantic rewrite (Table 2) and no word-level rewrite. However, new tests on 8 more rewrites show that, unlike laundering, each one is either detected or loses much of the capability.

  1. Word swaps keep the capability but are always detected (p < 0.001, even at the highest swap rate).
  2. The new semantic rewrites lose a third to a half of the capability, and the audit still detects most of them (18 of our 24 models at p < 0.05, against 0 of 6 laundered and 0 of 6 innocent models).
  3. Only laundering evades both the audit and SDR (Sec. 5). It also keeps more of the capability than the semantic rewrites.

Word swaps replace 10% to 70% of the content words at 4 rates with a masked language model (RoBERTa-base). Round-trip translation uses OPUS-MT, and the other rewrites use the paper's rewriter, Qwen2.5-7B-Instruct.

Training conditionAudit AUCCapability keptSDR AUC
Word swap0.781.010.77
Round-trip translation0.640.560.56
Paraphrase0.530.510.81
Simplification0.530.600.85
Summarize, then expand0.540.640.68
Register rewrite0.500.760.65
Laundered (ours)0.510.840.51

We added these results to Table 2 and App. E.8 (Table 15).

Natural fingerprints (W3)

We agree with the reviewer that natural fingerprints can survive regeneration. On images, an auditor who has D and reproduces our regeneration pipeline flags laundered models more often than innocent ones (78% vs 47% at the standard threshold, 47% vs 8% at the strict one). However, at a base rate of 1 in 100, about 95% of flagged models are false positives at the strict threshold, and 98% at the standard one, against 40% to 60% for the membership audit on direct training (Fig. 3). An audit of released models can face such a low base rate. On text, the fingerprint does not separate laundered models from innocent models trained with the same pipeline (22% of each flagged at the strict threshold).

Similar to the dataset classifier of Mansour & Heckel (2025), the fingerprint audit trains a classifier to distinguish the outputs of reference models trained on D and on public data. It flags a suspect model whose outputs score above all models trained on the same artist's other public works that the auditor never saw (standard threshold), or above all other innocent models (strict threshold). Each rate combines 36 measurements from 6 suspect models (2 artists, 3 seeds), each evaluated by classifiers trained on 3 seeds of the auditor's reference models, with 2 feature sets (CLIP embeddings, and the Fourier spectrum of each image's median-filter residual).

We added this audit and its results to Sec. 5 and App. G, with a sentence on how future audits could use natural fingerprints.

Please let us know whether these results address your three concerns.

Thank you for the additional experiments. However, in my understanding, two concerns still remain: the image study evaluates mostly isolated transformations rather than composed or adaptive regeneration pipelines, and several text rewrites substantially reduce capability, making it difficult to disentangle audit robustness from utility loss. Therefore, I maintain my original rating.

Reviewer_Y7MpOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes the concept of data laundering, where the licensee first fine-tunes a generator using licensed data, then generates new data from the generator, and trains and deploys the model using only the new data, thereby preserving some task capabilities while reducing the detection rate of post-hoc data traceability audits. The paper tests nine audit methods in image and text settings and proposes an attribution trade-off theorem for separable/inseparable signals.

Critical strengths

The paper identifies a significant data traceability threat with real-world implications and conducts systematic testing across multiple types of audits.

Critical weaknesses

The paper classifies signals into two categories: separable signals and inseparable signals. The former can be removed while preserving the target capability, while the latter is necessary to achieve the target capability. Based on this distinction, the paper derives the following theorem: the former can only be retained by regenerating the data, while the latter also appear in models that possess the same capability but do not use the licensed data. Therefore, neither can be reliably attributed to the licensed data. Currently, this reasoning relies heavily on the definitions of capability\textit{capability} and signal\textit{signal}. In particular, does the fact that “there exists a model that achieves the capability threshold without using the licensed data” suffice to infer that this model also possesses the specific signal measured by the audit? Different models may achieve the same capability through different internal representations. Thus, it is recommended to formalize the definitions of signal\textit{signal}, capability threshold\textit{capability threshold}, and attributability\textit{attributability}.

Questions for authors

Does the fact that “there exists a model that achieves the capability threshold without using the licensed data” suffice to infer that this model also possesses the specific signal measured by the audit?

2 comments
AuthorsOpenReview ↗

Response to Reviewer Y7Mp

We thank the reviewer for the insightful comments. We are committed to making all necessary changes and improvements. We formalized the notions the reviewer asked about and added the definitions to the revised paper (in blue).

Signal, capability threshold and attributability (W1)

A model that reaches the capability threshold without the licensed data need not carry the signal an audit measures. Theorem 1 assumes that it does only for inseparable signals, which by definition every capable model carries (App. C, condition (ii)). The theorem splits signals by this property. A signal that some capable model lacks is separable, and Proposition 1 bounds what an audit recovers from it about each record's membership by the best test on D′, without using BnegB_{\mathrm{neg}}.

We agree with the reviewer that the submitted Sec. 3.2 did not state these notions formally. We give formal definitions below, and the proof of Theorem 1 becomes a case split on them.

  1. Signal. An audit computes a statistic s(B)s(B) from the outputs or likelihoods of a model BB, where larger values indicate training on D, and flags BB when s(B)≥ts(B) \ge t. Internal representations therefore matter only through s(B)s(B).
  2. Capability threshold. U(B)U(B) is the fraction of a directly trained model's held-out capability that BB keeps (Sec. 2), and the capable class is Mτ={B:U(B)≥τ}M_\tau = \lbrace B : U(B) \ge \tau\rbrace, with τ\tau fixed before any experiment.
  3. Separable and inseparable. A signal is inseparable at tt if every B∈MτB \in M_\tau has s(B)≥ts(B) \ge t, and separable otherwise.
  4. Attributable. A signal is attributable to D at false-positive rate α\alpha if it flags at most a fraction α\alpha of the models in MτM_\tau trained without D, and a larger fraction of those trained on D′.

The inseparable branch then follows from condition (ii) by definition. BnegB_{\mathrm{neg}} is in MτM_\tau and every model in MτM_\tau is flagged, so the signal flags models trained without D and is not attributable at any α<1\alpha < 1. A signal carried by some capable models and not by others, as the reviewer describes, is separable, and so is the signal of every audit that flags fewer than all capable models trained without D. For these audits, the relevant part of Theorem 1 is the separable branch and its per-record bound (Proposition 1). Sec. 4.3 measures how much of the signal survives laundering.

Our experiments measure condition (ii) directly. On text, models trained on another author of the same period keep 0.92 and 0.93 of the capability gain (Sec. 4.2).

We added these definitions to Sec. 3.2, and App. C now presents the proof as this case split.

Please let us know whether these definitions address your concern.

Thank you for the clarification. Your response addresses my concerns, and I will retain my original score.

Reviewer_PSyHOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the robustness of training-data provenance audits against data laundering. The attacker fine-tunes a generator on a licensed dataset, samples a synthetic replacement dataset, and trains deployed models only on the synthetic data. Experiments investigate whether these models retain useful capabilities while evading membership, watermark, and similarity audits.

Critical strengths

The paper studies an important problem: whether data owners can reliably detect unauthorized use of their datasets. It offers an insightful observation that useful capabilities can survive data regeneration even when the signals used to trace their source become substantially weaker.

The evaluation covers multiple audit families, text and image settings, and several potential defenses. Direct-training positive controls and innocent-model controls help distinguish successful evasion from an audit that was ineffective to begin with.

The paper explicitly measures capability retention alongside detection, rather than treating degraded or unusable models as successful attacks. The from-scratch text models also provide useful control over pretraining contamination.

Critical weaknesses

The strongest evasion results depend on additional attacker resources that complicate their interpretation. Ordinary regeneration does not defeat DIAGNOSIS; successful evasion additionally uses knowledge of the watermark family, a specialized de-warping procedure, and 150 unmarked same-artist works, compared with 50 licensed images. A matched public-data-only baseline would help disentangle the contribution of laundering from watermark dilution and capability acquisition through these additional works.

The practical capability–cost–evasion tradeoff is not fully established. The paper reports generation costs and capability retention, but does not clearly demonstrate when laundering is preferable to obtaining comparable capabilities from alternative data. This leaves uncertainty about the practical benefit of laundering under the chosen capability metric. Comparisons at matched computational and data budgets, together with more source-specific capability measurements, would better establish what the provider gains and sacrifices.

Questions for authors

Please see weaknesses

2 comments
AuthorsOpenReview ↗

Response to Reviewer PSyH

We thank the reviewer for the insightful comments. We are committed to making all necessary changes and improvements. We performed the matched public-data baseline experiment the reviewer suggested and added the results to the revised paper (in blue).

Launderer's resources against DIAGNOSIS (W1)

We agree with the reviewer that the paper's strongest result against DIAGNOSIS also uses knowledge of the watermark family and 150 unmarked works of the same artist. However, new tests with the matched public-data baseline the reviewer suggested, and with dilution without laundering, show that the 150 works help only together with laundering, and that de-warping needs no dilution.

  1. Dilution without laundering does not remove the watermark. Training directly on the 50 marked and the 150 unmarked works triggers the detector on all 6 runs.
  2. De-warping defeats DIAGNOSIS without dilution. With the generator trained only on the 50 de-warped marked images, laundering meets the paper's pre-registered criterion (at most 1 in 3 runs trigger) on both settings, X1 and X2 (Table 13). Dilution without de-warping is enough on X2 (0 of 3 runs trigger) but not on X1 (2 of 3).
  3. Models trained on the 150 unmarked works alone never trigger the detector (0 of 6 runs).
Images usedLaunderer's stepsRuns rejected, X1Runs rejected, X2
50 markedDirect training1/11/1
150 unmarkedDirect training0/30/3
50 marked + 150 unmarkedDirect training3/33/3
50 markedLaundering3/33/3
50 marked + 150 unmarkedLaundering2/30/3
50 markedLaundering + de-warping1/61/3
50 marked + 150 unmarkedLaundering + de-warping0/60/3

We added the new rows to Table 13 and App. E.5, and the caption of Table 1 now marks the DIAGNOSIS entry as the one where laundering uses a step tailored to the audit (de-warping, Sec. 4.4).

Capability, cost and alternative data (W2)

Laundering is the cheaper option when the licensee already has D and obtaining comparable data costs more than one laundering pass, which takes 8.10 GPU-minutes (0.25 USD) on images and 15.26 GPU-minutes on text (App. B.3). Laundered models remain within 0.78% of direct training on FID and 0.82% on CLIP (App. B.4), and keep 0.85 and 0.82 of the capability gain on text (Sec. 4.2).

We agree with the reviewer that the paper did not compare laundering with alternative data at matched budgets. However, in our settings alternative data is free to obtain. In each setting, similar works are publicly available (other WikiArt paintings by the same painter, or books by a different author of the same period), and on text these reach most of the capability gain (Sec. 4.2). Outside our settings, comparable data can be costly.

We added this discussion to Sec. 7.

Please let us know whether these results address your two concerns.

Thank you for providing additional experiments and discussions. The rebuttal addresses most of my concerns, and I maintain the positive rating.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We thank the reviewers for the insightful comments, and are glad that they find the problem important (PSyH, Y7Mp, i82p) and our tests systematic across audit types, text and images, and defenses (PSyH, Y7Mp). We uploaded a revised paper with new and changed text in blue, and replied to each reviewer below. The revision adds the definitions and experiments the reviewers asked for.

  • Sec. 3.2 now gives formal definitions of an audit's signal, the capability threshold, separable signals and attributability, and App. C proves Theorem 1 as a case split on them (Y7Mp).
  • Sec. 4.3 and App. E.7 test in-place image transforms at 17 settings of rotation, resizing, color change and style transfer (i82p).
  • Table 2 and App. E.8 add word swaps at four rates and four semantic rewrites, scored by the audit and by SDR (i82p).
  • Sec. 5 and App. G add an audit of natural fingerprints on images and text (i82p).
  • Table 13 and App. E.5 add matched public-data baselines for DIAGNOSIS, and the caption of Table 1 now marks the step tailored to that audit (PSyH).
  • Sec. 7 discusses when laundering is the cheaper option for a licensee (PSyH).

Dear Reviewers,

The authors have replied to your initial comments. Please have a look and engage in the discussion.

Thanks!

Your AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.