acceptodds
Under review as a conference paper at ICLR 2027

Dataset Inference for Data Provenance and Privacy Auditing in Tabular Foundation Models

Abstract

Tabular foundation models (TFMs) are increasingly deployed through black-box APIs and trained on large collections of real-world tabular datasets. As this training paradigm becomes more widely adopted, private or proprietary datasets may be incorporated into pre-training corpora, yet there are currently no dedicated methods for determining whether a given tabular dataset was used to train a TFM. As a solution, we introduce the first dataset inference method for TFMs, aiming to infer whether a suspect dataset was part of a model’s pre-training data. We systematically analyze a broad collection of candidate signals that can be observed from a black-box TFM via input manipulation and find that individual signals provide only weak evidence of training data membership. However, by combining multiple complementary signals, we can infer dataset membership for several state-of-the-art TFMs trained on real tabular data, achieving up to ROCAUC dependent on the TFM. We then study factors that influence dataset identification, including pre-training data composition, model capacity, and the use of real vs. synthetic data. Our results show that our dataset inference method is a practical auditing tool for detecting privacy leakage and the use of proprietary datasets in TFMs.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
tabular foundation models, privacy, data provenance, auditing

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ZN7NOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces the first dataset inference method for tabular foundation models (TFMs). It asks whether a whole suspect dataset was in a black-box TFM’s pre-training corpus. Individual classical signals like loss are weak. The authors instead build dataset-level fingerprints from 1,021 features, formed from 19 input manipulations, 14 output signals and 11 reductions. They use an Isolation Forest for external audits, where only non-members are known, and a trained classifier for internal audits, where member labels are known. Each result is compared against a blind baseline TFM that never saw the data.

On four TFMs, the internal audit reaches up to 0.988 ROCAUC on SAP-RPT-1-OSS and about 0.70 on TabDPT. nanoTabPFN and Real-TabPFN-2.5 show no meaningful leakage, and external-audit gains are weaker.

Critical strengths
  • The paper tackles an important problem. Detecting the use of proprietary data in model training is critical for preventing unauthorized data use and data leakage.
  • Manipulating the input to probe the model’s knowledge is an interesting idea.
  • The evaluation is comprehensive. Beyond reporting AUCs, the experiments explore the most revealing features and the factors that affect robustness to membership leakage.
Critical weaknesses
  • Performance is weak. In Table 1, ROCAUC is close to random guessing. In Figure 6, even the internal audit shows strong separation in only one setting.
  • For a black-box model, the paper assumes access to the output distribution, which is often not available.
  • Section 4.1 is unclear about how the loss is computed, and about which examples are in context and which are queries. This needs to be made clear before the section is presented.
  • Before introducing the manipulations, the paper should introduce the base case: how datasets get combined and split, and which part is in context and which serves as queries.
Questions for authors
  • The setting detects whether an entire dataset was used. What if the dataset was sampled?
  • Why must each dataset be treated as a single unit? Would further sampling and computation provide more signal?
  • How exactly is the loss computed in Section 4.1, and what serves as the in-context examples versus the query data points?
4 comments

We thank the reviewer for their feedback and for recognizing that our work “tackles an important problem,” that our approach to using manipulations is “an interesting idea,” and that our evaluation is “comprehensive.”

W1: Performance is weak. In Table 1, ROCAUC is close to random guessing

Low AUC does not imply the method is weak, since some TFMs may simply not memorize membership information. In such a case, no DI method would do better than random guessing.

We investigated the degree of memorization, and performance of our DI method throughout Sec. 5.2 and 5.4. Table 1 shows the AUC for the external auditor varying from random guessing (0.498 on nanoTabPFN) to meaningful performance (0.652 on SAP-RPT-1-OSS). Table 2 shows that the more powerful auditor achieves almost perfect success on SAP-RPT-1-OSS, but still no success on Real-TabPFN-2.5 and nanoTabPFN and only modest success on TabDPT. These results suggest that our method is highly performant in identifying membership signals when they exist, but that some TFMs simply memorize less.

Moreover, Table 1 shows a DI setting that is strictly more challenging than Table 2, and accordingly performance levels are higher in Table 2.

W2: the paper assumes access to the output distribution, which is often not available.

Unlike LLMs which often only provide completions not logits, for commercial TFMs the full output distribution is always provided at present [1-3]. Quantification of uncertainty for model predictions is seen as a critical feature for tabular predictive decisioning, so it is realistic to assume an auditor will have access to it.

Q1: The setting detects whether an entire dataset was used. What if the dataset was sampled?

If two datasets share some rows, but not all, they are not the same dataset. Hence, a sampled version of a training member dataset is by definition a non-member. This is why we labeled only the full datasets as members in our setting. If a sampled dataset were tested, it would need to be labeled as a non-member. Overall, it is a nuanced question: “how large must modifications to a member dataset be before we expect dataset inference to treat it as a non-member?” There is no single correct answer. This is closely linked to the next question.

Q2: Why must each dataset be treated as a single unit?

A TFM performs predictions on query samples using context examples with an ICL mechanism. During training, optimization is informed by query and context samples which come from the same dataset. Unlike LLMs, which can autoregressively output data with no or little input data, TFMs are inherently trained on this combination of context and query data from one dataset at a time. Therefore, treating each dataset as the membership unit is a natural conclusion from how they are being trained, as we outline in Sec. 2.2.

Q2 (cont’d): Would further sampling and computation provide more signal?

For dataset inference, further sampling or computation to modify a dataset would change its identity, and may flip the label from member to non-member. However, we note that in the process of creating membership signals, we use many controlled dataset manipulations to study the reaction of the TFM, including subsampling (Table 5 in App C), and these do increase the signal. As every input manipulation requires an additional forward pass, this is an increase in compute. To isolate the effect of further computation, we test using a fixed set of fingerprints and sample the target model on one random query/context split (k=1, as in the paper shows) and on multiple splits (k>1). As this requires fitting and extracting more data, multiplying the compute cost, we run this ablation on a balanced subset of 1,000 tables for models with more datasets (SAP-RPT-1-OSS, nanoTabPFN). Additional sampling does not significantly change the target’s ROC-AUC for any model, indicating that more sampling without additional signals does not provide more signal beyond our proposed method.

Internal audit (TabPFN-2.5 classifier): ROC-AUC, target (blind)

Target TFMTablesk = 1 (paper)k = 2k = 3k = 5
SAP-RPT-1-OSS10000.987 (0.699)0.988 (0.669)0.991 (0.685)0.991 (0.679)
TabDPT2240.644 (0.583)0.662 (0.645)0.680 (0.632)0.662 (0.688)
nanoTabPFN10000.501 (0.488)0.506 (0.517)0.485 (0.463)0.485 (0.506)
Real-TabPFN-2.5700.468 (0.406)0.448 (0.464)0.527 (0.541)0.536 (0.500)

External audit (Isolation Forest): ROC-AUC, target (blind)

Target TFMTablesk = 1 (paper)k = 2k = 3k = 5
SAP-RPT-1-OSS10000.769 (0.413)0.761 (0.387)0.772 (0.393)0.775 (0.414)
TabDPT2240.566 (0.505)0.569 (0.510)0.564 (0.507)0.545 (0.500)
nanoTabPFN10000.493 (0.495)0.499 (0.496)0.503 (0.498)0.496 (0.478)
Real-TabPFN-2.5700.499 (0.529)0.510 (0.533)0.509 (0.535)0.584 (0.608)

Q3: How exactly is the loss computed in Section 4.1?

A given dataset is divided into a context and query split as in Figure 1, and $y_{ctx}$ is masked before it is fed into the TFM. The loss is computed between labels $y_{ctx}$ and predictions $\hat y_{ctx}$, averaged over the context rows.

Figure 2 in Section 4.1 uses the average cross entropy loss to illustrate that member and non-member datasets are not distinguishable by loss. Loss as a standalone feature tracks dataset difficulty rather than membership. For our dataset inference method, we incorporate more signals, including different loss functions, such as KL divergence or mean squared error, as Table 4 outlines.

Q3 (cont’d): What serves as the in-context examples versus the query data points?

This depends on the manipulation at hand, which Table 5 describes in detail. By default, each dataset is split disjointly into query and context (Figure 1). Some manipulations affect context or query in isolation. E.g., the row sampling manipulation (M4) removes context rows to measure how robust the prediction is to having fewer ICL examples. Other manipulations change how the context and query samples are composed. E.g., the “Query = Context” manipulation (M2) sets both sets to the same samples. The intuition is that memorization might make a model profit less from this contamination, while a model never trained on the data would show a higher increase in utility.

We hope the above answers address the Reviewer's concerns and would greatly appreciate it if they would increase their score accordingly.

References

[1]: Predictive distribution. (n.d.). Prior Labs. Retrieved November 9, 2026, from https://docs.priorlabs.ai/capabilities/predictive-distribution
[2]: Understand prediction results. (n.d.). NVIDIA Kumo. Retrieved November 9, 2026, from https://docs.nvidia.com/sdgm/client/kumo-relational/user-guide/work-with-prediction-results/prediction-results
[3]: TabH2O — Tabular prediction API. (n.d.). tabh2o.h2oai.com. Retrieved November 9, 2026, from https://tabh2o.h2oai.com/docs

Thank you for the response. My concern about assuming access to the output distribution has been addressed.

However, I am still not convinced that the full dataset must be treated as a single unit. If an adversary can bypass detection by subsampling or transforming the data, the method has limited practical value.

I also remain concerned about the evaluation. Strong results are limited to SAP-RPT-1-OSS in the internal-audit setting, and the external-audit gains are weak, so it is unclear whether the method supports meaningful real-world applications.

I will keep my current score.

We sincerely thank the Reviewer for the follow-up and are glad the concern about output distribution access has been resolved.

full dataset must be treated as a single unit
bypass detection by subsampling or transforming the data

TFM training already transforms and subsamples each table. SAP-RPT-1-OSS, for example, picks a prediction column at random and subsamples each table to at most 1,000 rows. Our main results are obtained on models trained this way, which already shows that such transformations do not bypass detection. We additionally test subsampling specifically, by auditing with tables of increasing size, so that a shrinking share of the auditor's rows was used in training. We report target values with the blind baseline in parentheses:

Auditor's table row countShare of rows used in trainingROC-AUCTPR at 1% FPRTPR at 10% FPR
250100%0.850 (0.496)0.291 (0.037)0.622 (0.121)
500100%0.918 (0.495)0.395 (0.031)0.773 (0.131)
1,000100%0.909 (0.418)0.335 (0.001)0.748 (0.058)
1,10091%0.907 (0.428)0.279 (0.006)0.746 (0.060)
1,20083%0.890 (0.425)0.230 (0.005)0.686 (0.057)
1,50067%0.847 (0.420)0.148 (0.005)0.583 (0.054)
2,00050%0.793 (0.380)0.047 (0.005)0.341 (0.031)
4,00025%0.840 (0.370)0.067 (0.014)0.393 (0.041)

Subsampling weakens detection but does not bypass it. When all of the auditor's rows were used in training, the audit reaches 0.85 to 0.92 ROC-AUC. When only 25% to 50% were used, it still reaches 0.79 to 0.85. The blind baseline sits at or below chance throughout (0.37 to 0.50), so on a model that never saw the data, the same shift between members and non-members points away from membership and cannot explain the target's result. At 1% FPR, the TPR falls from 0.29–0.40 to 0.05–0.08, which remains several times the blind baseline (at most 0.014 in these rows). An adversary would therefore have to withhold most of a dataset to reduce detection, and a clear signal remains even then. Adversarial transformations are not covered by this experiment, and we acknowledge this as a limitation.

Strong results are limited to SAP-RPT-1-OSS in the internal-audit setting, and the external-audit gains are weak

Varying results are not a weakness of the method but due to differences in the models. A TFM that does not retain membership information will show weak dataset inference under any method. On SAP-RPT-1-OSS, the internal audit detects 92% of member datasets at 1% FPR, and the external audit, without any knowledge of the training data, still flags 512 member datasets at 0.5% FPR. On nanoTabPFN and Real-TabPFN-2.5, neither audit finds a signal even with labeled members, and on TabDPT only a modest one, which suggests that these models retain far less membership information.

meaningful real-world applications

Our auditing procedures have already been used in the real world to uncover cases of suspected cheating on the most widely used TFM leaderboard - TabArena [1]. The maintainers of TabArena (who are independent from us) require TFMs to not be trained on the TabArena test set in order to be ranked on the leaderboard. They have tested submitted TFMs using our internal auditing procedure, and found one submitted TFM (LimiX-2 [2]) flagged datasets from the test set as training members - a clear violation of the rules. This led to LimiX-2 being removed from the leaderboard and replaced. This shows that our methods are already practical and support transparent research conduct, similar to the motivations we presented about data owners verifying how their datasets have been used.

These findings were disclosed in a blog post by the TabArena authors. Since the blog post references our work directly, we cannot link it without violating anonymity.

[1] Erickson et al. “TabArena: A Living Benchmark for Machine Learning on Tabular Data” NeurIPS 2025

[2] Zhang et al. “LimiX-2: A Contextual Mechanism Network Towards General Structured-Data Intelligence” arXiv:2609.17488

Reviewer_uC5VOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a black-box dataset inference method for data provenance and privacy auditing in Tabular Foundation Models (TFMs), aiming to determine whether a suspect dataset was used during model pre-training. The proposed method constructs dataset-level fingerprints by combining multiple complementary signals for membership inference. The authors consider both external auditing, which uses an Isolation Forest trained on known non-member datasets, and internal auditing, which employs supervised classifiers with known member and non-member datasets. Experiments on four TFMs achieve up to 0.988 ROC-AUC in the internal auditing setting. The paper further investigates the effects of training data composition, model capacity, and real versus synthetic data on dataset membership leakage.

Critical strengths
  1. The paper systematically investigates dataset-level membership inference for TFMs, which is an interesting privacy auditing problem. The proposed framework explores diverse input manipulations and output signals to construct dataset fingerprints.
  2. The experiments cover four different TFMs with different pre-training strategies, including real and synthetic datasets.
  3. The paper is generally well-written and easy to follow.
Critical weaknesses
  1. Dataset inference has already been extensively studied in previous works. The proposed method mainly extends existing dataset inference techniques to TFMs by combining multiple membership signals and input manipulations. The fundamental idea of aggregating multiple signals for dataset-level membership inference is not new, and the paper does not sufficiently demonstrate its methodological novelty over prior approaches.
  2. All evaluated models use publicly released weights or are trained locally by the authors. Although black-box access can be simulated using these models, the experiments do not demonstrate the feasibility of the proposed attack against actual commercial TFM APIs. Therefore, the practical applicability of the proposed method under realistic deployment constraints remains unclear.
  3. The proposed method assumes that an attacker can obtain the full predictive probability distribution from the target TFM. However, not all commercial black-box APIs necessarily expose such detailed outputs. Several important membership signals, including confidence, cross-entropy, entropy, and KL divergence, depend on access to predictive probabilities. The paper does not evaluate whether the proposed method remains effective under a more restrictive label-only black-box setting, which limits the generality of its threat model.
  4. The paper primarily compares its method with individual membership signals, such as average loss and confidence, and introduces a blind baseline to account for distributional differences. However, the blind baseline is not a competing attack method. The paper lacks systematic comparisons with existing dataset inference methods or their adaptations to TFMs. Without such comparisons, it is difficult to determine whether the proposed method provides substantial improvements over prior techniques.
  5. The external auditing scenario is particularly important because it aligns with the paper's motivation of identifying unauthorized use of proprietary datasets through black-box APIs. However, the experimental results are relatively weak, with ROC-AUC scores of only 0.652, 0.619, 0.498, and 0.565 across the four evaluated TFMs. In contrast, the highest reported ROC-AUC of 0.988 is achieved in the internal auditing setting, which assumes access to known member and non-member datasets. These results raise concerns about the practical effectiveness of the proposed method for external data owners. Furthermore, the paper does not report low-FPR metrics such as TPR@1%FPR or TPR@5%FPR, which are important for evaluating the reliability of privacy auditing methods.
Questions for authors

See Weaknesses above.

4 comments

W1: The proposed method mainly extends existing dataset inference techniques to TFMs by combining multiple membership signals and input manipulations

The main novelty of our method is to consider the TFM ICL mechanism. No existing DI method is applicable to the ICL predictions inherent to TFMs, and they perform poorly when naively applied as we showed in Section 4.1. We construct membership inference signals out of various input manipulations, not just as a combination of signals like previous membership inference literature did, but because it inherently captures the uniquely different architecture. These input manipulations are novel.

W2: feasibility of the proposed attack against actual commercial TFM APIs and practical applicability of the proposed method under realistic deployment constraints

Commercial TFM APIs occupy a different space than LLM APIs, and we should not conflate the two. LLM APIs are consumer-facing; TFM APIs are currently business-facing, and as such are not directly accessible to researchers. However, some of the same TFMs used commercially are available as open-weight models for academic purposes. These are the models we used, and we emulate the conditions of a black-box API in our external auditor setup. The external auditor only holds non-member data (possibly out-of-distribution), has black-box access to the model, and has no information about the model architecture, so they cannot train shadow models. Therefore, the external auditor scenario closely resembles realistic deployment constraints.

W3: proposed method assumes that an attacker can obtain the full predictive probability distribution

Unlike LLMs which often only provide completions not logits, for commercial TFMs the full output distribution is always provided at present [1-3]. Quantification of uncertainty for model predictions is seen as a critical feature for tabular predictive decisioning, so it is realistic to assume an auditor will have access to it.

W4: systematic comparisons with existing dataset inference methods or their adaptations to TFMs

Our work is the first dataset inference method for TFMs. As such, there is no competitive baseline to compare to. In Section 4.1 we showed that generic DI methods which are not designed specifically for TFMs perform poorly when naively applied to TFMs.

W5: TPR at low FPR results

We have added TPR at 1%, 5%, and 10% FPR for both audit settings to the paper in App. D.2 and D.3.

The external audit is the hardest setting. On SAP-RPT-1-OSS the external auditor achieves 4% TPR@1%FPR on the target model (and 1% TPR@1%FPR on the blind baseline). At an FPR of less than 0.1%, they still achieve a TPR of 2.8%, identifying roughly 450 member tables with two false positives. For larger FPRs the TPR regresses to the FPR level, so the membership leakage is concentrated in a minority of strongly exposed and confidently detected datasets.

If the external auditor wishes to make high confidence claims, they can adopt the conservative Benjamini-Hochberg step-up procedure [4] at a 5% FDR level. For SAP-RPT-1-OSS, the external auditor flags 512 member datasets and achieves a 3.2% TPR at 0.5% FPR. For the other models, no datasets get flagged, as the signal is not decisive enough, so the auditor would make no false accusations.

Target TFMTrue positives (blind)TPR (blind)FPR (blind)
SAP-RPT-1-OSS512 (0)3.2% (0%)0.5% (-)
TabDPT0 (0)0% (0%)n/a (-)
nanoTabPFN0 (0)0% (0%)n/a (-)
Real-TabPFN-2.50 (0)0% (0%)n/a (-)

In the easier internal audit setting, the auditor achieves 92% TPR@1%FPR on SAP-RPT-1-OSS. The performance for the blind baseline is 17% TPR@1%FPR, confirming a significant distribution shift between member and non-member datasets but highlighting the much higher performance on the actual target model. On TabDPT, the internal auditor achieves 4% TPR@1%FPR with the blind baseline at 1% TPR@1%FPR.

Target TPR (blind TPR)0% FPR1% FPR5% FPR10% FPR
SAP-RPT-1-OSS0.740 (0.040)0.916 (0.166)0.972 (0.265)0.987 (0.276)
TabDPT0.043 (0.000)0.043 (0.009)0.103 (0.094)0.214 (0.162)
nanoTabPFN0.000 (0.000)0.004 (0.007)0.055 (0.029)0.114 (0.029)
Real-TabPFN-2.50.000 (0.000)0.000 (0.000)0.053 (0.053)0.158 (0.105)

We hope the above answers address the Reviewer's concerns and would greatly appreciate it if they would increase their score accordingly.

AuthorsOpenReview ↗

References

[1]: Predictive distribution. (n.d.). Prior Labs. Retrieved November 9, 2026, from https://docs.priorlabs.ai/capabilities/predictive-distribution
[2]: Understand prediction results. (n.d.). NVIDIA Kumo. Retrieved November 9, 2026, from https://docs.nvidia.com/sdgm/client/kumo-relational/user-guide/work-with-prediction-results/prediction-results
[3]: TabH2O — Tabular prediction API. (n.d.). tabh2o.h2oai.com. Retrieved November 9, 2026, from https://tabh2o.h2oai.com/docs
[4]: Benjamini, Y., & Hochberg, Y. (1995). Controlling the false discovery rate: a practical and powerful approach to multiple testing. Journal of the Royal statistical society: series B (Methodological), 57(1), 289-300.

Thank you for the response. I appreciate the authors' efforts to address my concerns.

However, my main concerns regarding methodological novelty, the lack of systematic comparisons with existing dataset inference methods, and the limited effectiveness of external auditing remain insufficiently addressed. Therefore, I will maintain my original assessment.

We thank the reviewer for their engagement.

methodological novelty, the lack of systematic comparisons with existing dataset inference methods

We propose the first method for investigating TFM pretraining dataset inference. As such, there is no existing method that applies to the unique ICL architecture of TFMs. Still, we added adaptations of existing methods [5-6], which compute membership signals from model outputs on unmodified suspect samples, aggregate them, and compare them against reference data. We adapt them by computing the accuracy, loss, correct-class confidence, maximum confidence, prediction margin, and entropy for each table using the unmodified context.

Internal audit (TabPFN-2.5 classifier): ROC-AUC, target (blind)

Target TFMTablesStandard signalsOur method
SAP-RPT-1-OSS20,1000.857 (0.852)0.988 (0.689)
TabDPT2240.513 (0.450)0.695 (0.611)
nanoTabPFN20,0000.502 (0.500)0.539 (0.497)
Real-TabPFN-2.5700.719 (0.727)0.542 (0.560)

Internal audit (TabPFN-2.5 classifier): TPR at 1% FPR, target (blind)

Target TFMStandard signalsOur method
SAP-RPT-1-OSS0.238 (0.251)0.916 (0.166)
TabDPT0.000 (0.000)0.043 (0.009)
nanoTabPFN0.006 (0.007)0.004 (0.007)
Real-TabPFN-2.50.000 (0.053)0.000 (0.000)

External audit: ROC-AUC, target (blind)

Target TFMStandard signalsOur method
SAP-RPT-1-OSS0.422 (0.443)0.652 (0.442)
TabDPT0.544 (0.456)0.583 (0.472)
nanoTabPFN0.504 (0.507)0.502 (0.502)
Real-TabPFN-2.50.629 (0.628)0.566 (0.581)

The results show that our proposed approach outperforms adaptations of existing methods to TFMs. For almost all tests, the standard signals have lower performance and do not perform better on the target model than the blind baseline, a strong sign that they capture distribution shifts more than membership signals. Where standard signals score higher than our method, on Real-TabPFN-2.5 (0.719 internal, 0.629 external), the blind baseline matches them (0.727 and 0.628), so these values reflect distribution shift, not membership. For example, an internal audit with standard signals on SAP-RPT-1-OSS achieves a TPR@1%FPR of 23.8%, but when computed on a blind baseline, which has not been trained on any of the tested tables, it also achieves 25.1% TPR@1%FPR. Our method, in contrast, clearly exceeds its blind baseline on SAP-RPT-1-OSS and TabDPT, the two models that leak membership information. On SAP-RPT-1-OSS, it reaches 91.6% TPR@1%FPR against 16.6% for the blind baseline. A blind baseline above chance indicates distribution shift, so what isolates membership is the excess of the target over the blind baseline. On SAP-RPT-1-OSS, the standard signals show none, whereas our method gains 75 percentage points in TPR@1%FPR. This gain comes from our input manipulations and signal selection, and it is the methodological contribution of our work.

limited effectiveness of external auditing

We agree that the external auditing remains challenging, especially under the black-box setup we use. Still, our low FPR results show that on SAP-RPT-1-OSS, our method flags 512 member datasets at 0.5% FPR without any knowledge of the training data or about the model. On the three models with a faint membership signal, it flags no datasets and thus makes no false accusations. The adapted existing methods, by contrast, do not exceed their blind baseline in the external audit on SAP-RPT-1-OSS (0.422 vs. 0.443 ROC-AUC). Our method is therefore the first to give a data owner with only black-box access a high-confidence provenance signal on a TFM.

[5]: Maini, P., Yaghini, M., & Papernot, N. (2021). Dataset inference: Ownership resolution in machine learning. arXiv preprint arXiv:2104.10706.

[6]: Maini, P., Jia, H., Papernot, N., & Dziedzic, A. (2024). LLM Dataset Inference: Did you train on my dataset?. Advances in Neural Information Processing Systems, 37, 124069-124092.

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

The paper proposes black-box dataset inference for tabular foundation models, constructing fingerprints from input manipulations and predictive signals. It distinguishes external auditing via anomaly detection from internal auditing using labeled member/nonmember datasets.

Critical strengths

The problem is important, and the probes are thoughtfully adapted to the in-context prediction interface. Combining signals yields strong internal-audit discrimination on SAP-RPT-1-OSS, reaching 0.988 AUC. The inclusion of blind baselines is valuable because their substantial predictive performance exposes distributional confounding that could otherwise go unnoticed.

Critical weaknesses

The strongest result is potentially confounded by table size. Appendix D.1 defines SAP-RPT-1-OSS members as tables with more than 150 rows and nonmembers as tables with at most 150 rows. Although size-related features are excluded, model responses may still indirectly encode size. Comparing against a different blind model does not establish that the excess AUC specifically measures membership leakage: the models may respond differently to the same distribution shift. Size/source-matched controls, ideally with randomized training membership, are needed.

The external-audit conformal calibration is not justified as written. Section 4.2 appears to reuse known nonmember datasets for fitting the Isolation Forest and calibrating its scores, while feature selection depends on suspect-versus-known separation. No independent calibration split or exchangeability argument is provided. Consequently, the stated procedure does not establish valid false-positive control. The authors should clarify the splitting protocol and demonstrate calibration after the complete feature-selection pipeline.

Practical auditing claims exceed the demonstrated external performance. Table 1 reports external-audit AUCs of only 0.652 for SAP-RPT-1-OSS and 0.619 for TabDPT, substantially below the headline internal-audit result. Practical provenance claims require detection power at low false-positive rates, sensitivity to reference-dataset choice, and query-cost measurements. Additionally, failure of this particular attack should not be interpreted as evidence that a model retains little membership information.

Questions for authors

Does the main result persist when member and nonmember tables have matched sizes and sources? Are feature selection, anomaly-model fitting, and calibration performed on disjoint datasets, and what false-positive rates are observed under a fully nonmember evaluation?

2 comments

We thank the reviewer for the constructive feedback, and for positively highlighting that the problem we study is “important”, that our method is “thoughtfully adapted”, and our “inclusion of blind baselines is valuable”.

W1/Q1: Does the main result persist when member and nonmember tables have matched sizes and sources?

We run an additional experiment where we investigate datasets with the same number of rows on SAP-RPT-OSS-1, where size determines membership. To do this, we randomly select 1,200 member tables and 1,200 non-member tables with at least 64 rows. We randomly subsample all tables to 64 rows. Performing the internal audit on this subset gives a target ROC-AUC of 0.754 (blind 0.666). This indicates there is a membership signal after size-matching. The signal is lower though, which is probably due to the significant subsampling of the tables and the fact that the model was not actually trained on such small tables. This setting is somewhat contrived, and as such, we expect the main results presented in Section 5 to better reflect real scenarios.

W2/Q2: Are feature selection, anomaly-model fitting, and calibration performed on disjoint datasets

We originally calibrated on the same known datasets used to fit the IsolationForest because two of our held-out pools are small (54 and 26 known non-members for TabDPT and Real-TabPFN-2.5). We change the procedure to split the held-out data, with one half being used to fit the iForest and the other half used for calibration. For SAP-RPT-1-OSS, where 1,960 known non-member datasets are available, the results are unchanged (ROC-AUC 0.67 target/0.44 blind, realized FPR of 1.0/5.0/10.1% at nominal 1/5/10% calibration). For nanoTabPFN (5,000 known datasets), the audit remains at chance (ROC-AUC 0.50 / 0.50) and is conservative (realized FPR 0.7 / 3.8 / 7.5%). For TabDPT and Real-TabPFN-2.5, the split leaves only 27 and 13 calibration datasets, so the lowest p-values are around 4% and 7%. The audit performance is slightly lower (ROC-AUC 0.56/0.50 for TabDPT, 0.53/0.46 for Real-TabPFN-2.5), and FPRs are conservative (6.0% and 5.3% at nominal 10%).

W3/Q3: what false-positive rates are observed under a fully nonmember evaluation?

We evaluate the FPR when none of the datasets in the suspect set of the external auditor are members using the procedure from Sec. 4.2, and thresholding at a chosen p-value to nominally control errors. All realized FPRs are on average at or below the nominal value. For tiny fitting and calibration pools, such as Real-TabPFN-2.5, which uses 12 datasets to fit the IsolationForest and 13 to calibrate, the variance is high, with a realized TPR of 8.0 ± 10.8% at 10% nominal. We have added this to App. D.2.

Realized FPRs for external audit with split conformal p-values

ModelSuspectsRealised FPR at nominal 1%at 5%at 10%
SAP-RPT-1-OSS1,9610.9 ± 0.2%4.2 ± 0.4%9.1 ± 0.9%
nanoTabPFN5,0001.1 ± 0.2%5.0 ± 0.3%10.1 ± 0.8%
TabDPT530.0 ± 0.0%4.5 ± 6.2%6.0 ± 5.7%
Real-TabPFN-2.5250.0 ± 0.0%0.0 ± 0.0%8.0 ± 10.8%

W3 (cont’d): external-audit AUCs [...] substantially below the headline internal-audit result

This is not a weakness. The external audit setting is strictly harder than the internal audit, so it is completely expected that external-audit AUCs are below internal-audit AUCs. This result serves as confirmation that more accurate dataset inference is possible when the internal auditor leverages additional information.

W3 (cont’d): Query-cost to the external auditor

TFMs are much smaller models than, for example, LLMs. As such, inference cost is of lower concern for TFMs. Parameter counts for the models we used are as follows: SAP-RPT-1-OSS = 116M, TabDPT = 77M, RealTabPFN-2.5 = 10M, and nanoTabPFN = 3.7M. Also, these models are not hosted as proprietary models like LLMs, so there is no standard quantification of token cost in dollars.

W3 (cont’d): sensitivity to reference-dataset choice

Each of the four models we tested uses a different set of reference datasets ($D_{known}$), because their training members and non-members naturally differ. These details were given in Appendix D.1. Hence, the results in Section 5.2 incorporate different reference dataset choices.

W3 (cont’d): TPR at low FPR results

We have added TPR at 1%, 5%, and 10% FPR for both audit settings to the paper in App. D.2 and D.3.

The external audit is the hardest setting. On SAP-RPT-1-OSS the external auditor achieves 4% TPR@1%FPR on the target model (and 1% TPR@1%FPR on the blind baseline). At an FPR of less than 0.1%, they still achieve a TPR of 2.8%, identifying roughly 450 member tables with two false positives. For larger FPRs the TPR regresses to the FPR level, so the membership leakage is concentrated in a minority of strongly exposed and confidently detected datasets.

If the external auditor wishes to make high confidence claims, they can adopt the conservative Benjamini-Hochberg step-up procedure [1] at a 5% FDR level. For SAP-RPT-1-OSS, the external auditor flags 512 member datasets and achieves a 3.2% TPR at 0.5% FPR. For the other models, no datasets get flagged, as the signal is not decisive enough, so the auditor would make no false accusations.

Target TFMTrue positives (blind)TPR (blind)FPR (blind)
SAP-RPT-1-OSS512 (0)3.2% (0%)0.5% (-)
TabDPT0 (0)0% (0%)n/a (-)
nanoTabPFN0 (0)0% (0%)n/a (-)
Real-TabPFN-2.50 (0)0% (0%)n/a (-)

In the easier internal audit setting, the auditor achieves 92% TPR@1%FPR on SAP-RPT-1-OSS. The performance for the blind baseline is 17% TPR@1%FPR, confirming a significant distribution shift between member and non-member datasets but highlighting the much higher performance on the actual target model. On TabDPT, the internal auditor achieves 4% TPR@1%FPR with the blind baseline at 1% TPR@1%FPR.

Target TPR (blind TPR)0% FPR1% FPR5% FPR10% FPR
SAP-RPT-1-OSS0.740 (0.040)0.916 (0.166)0.972 (0.265)0.987 (0.276)
TabDPT0.043 (0.000)0.043 (0.009)0.103 (0.094)0.214 (0.162)
nanoTabPFN0.000 (0.000)0.004 (0.007)0.055 (0.029)0.114 (0.029)
Real-TabPFN-2.50.000 (0.000)0.000 (0.000)0.053 (0.053)0.158 (0.105)

W3 (cont’d): should not be interpreted as evidence that a model retains little membership information

This is incorrect. Failure of our proposed attack is some evidence that a model retains little membership information, albeit not complete and convincing evidence by itself. No other research has proposed methods to check whether TFMs retain membership information about their training data, whereas our results contribute evidence as to the varying levels that these models retain such information. We probed models using 678 membership signals, and for nanoTabPFN, Real-TabPFN-2.5, and TabDPT, even for the internal auditor only a modest success over the blind baseline was observed, while for SAP-RPT-1-OSS, a significant difference emerged.

We hope the above answers address the Reviewer's concerns and would greatly appreciate it if they would increase their score accordingly.

References

[1]: Benjamini, Y., & Hochberg, Y. (1995). Controlling the false discovery rate: a practical and powerful approach to multiple testing. Journal of the Royal statistical society: series B (Methodological), 57(1), 289-300.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all reviewers for their detailed and constructive feedback. As the discussion period is coming to a close, we would like to summarize the main changes and new results, several of which address concerns shared across reviews:

  • Detection at low false-positive rates (SzBS, uC5V): We added TPR at 1%, 5%, and 10% FPR for both audit settings (App. D.2 and D.3). The internal auditor reaches 92% TPR at 1% FPR on SAP-RPT-1-OSS (17% blind). For the external auditor, we additionally evaluate a new decision rule for high-confidence claims, the Benjamini-Hochberg step-up procedure [1]. It flags 512 member datasets at 0.5% FPR on SAP-RPT-1-OSS and flags nothing on the other three models. The method can therefore support high-confidence provenance claims where leakage exists, while avoiding false claims.
  • Access to the output distribution (uC5V, ZN7N): Current commercial TFM APIs expose the full predictive distribution [2-4], because uncertainty quantification is a core feature for tabular prediction. This is notably different from LLM APIs. Our threat model therefore matches how TFMs are deployed today.
  • External-audit performance and low AUCs (SzBS, uC5V, ZN7N): The external audit is strictly harder than the internal audit, so lower AUCs are expected. Even with labeled members, the internal auditor finds almost no signal on nanoTabPFN and Real-TabPFN-2.5 and only a modest one on TabDPT, while it reaches 0.988 on SAP-RPT-1-OSS. Low AUCs on some models likely reflects how much membership information each model retains rather than a limitation of the method.
  • Size confound (SzBS): We repeated the internal audit on SAP-RPT-1-OSS with 1,200 member and 1,200 non-member tables, all subsampled to 64 rows. The target model reaches 0.754 ROC-AUC (0.666 blind), showing that a membership signal remains strong once table size is matched. The signal is weaker than in our main results, which is expected as the heavy subsampling removes most of each table, and the model was never trained on such small tables, which makes this test less realistic than what we did in Section 5.
  • Validity of the external-audit calibration (SzBS): We fit the Isolation Forest and calibrate the conformal p-values on disjoint halves of the known non-member data, and we evaluate on suspect sets containing no members. The audit results change little under this split. Realized false-positive rates are close to or below the nominal levels for all four models, so the stated error control holds empirically.
  • Novelty and baselines (uC5V): No existing dataset inference method applies to the in-context prediction mechanism of TFMs, and generic signals perform poorly (Sec. 4.1). Our input manipulations are designed specifically for this mechanism, which is what makes dataset inference for TFMs possible.
  • Additional sampling (ZN7N): We repeated both audits with up to five random context/query splits per dataset. ROC-AUC does not change substantially for any model, indicating that the signal comes from the diversity of our membership fingerprints rather than from repeated sampling.
  • Clarity (ZN7N): We clarified how the loss in Sec. 4.1 is computed, and how each dataset is divided into context and query samples, both in the base case and under each manipulation.

We would be grateful if the reviewers could let us know whether the responses and revisions address their concerns. If so, we kindly ask the reviewers to consider updating their scores accordingly. If any questions remain, we are happy to clarify them in the remaining time.

References

[1]: Benjamini, Y., & Hochberg, Y. (1995). Controlling the false discovery rate: a practical and powerful approach to multiple testing. Journal of the Royal Statistical Society: Series B (Methodological), 57(1), 289-300.
[2]: Predictive distribution. (n.d.). Prior Labs. Retrieved November 9, 2026, from https://docs.priorlabs.ai/capabilities/predictive-distribution
[3]: Understand prediction results. (n.d.). NVIDIA Kumo. Retrieved November 9, 2026, from https://docs.nvidia.com/sdgm/client/kumo-relational/user-guide/work-with-prediction-results/prediction-results
[4]: TabH2O — Tabular prediction API. (n.d.). tabh2o.h2oai.com. Retrieved November 9, 2026, from https://tabh2o.h2oai.com/docs

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.