- 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
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 count | Share of rows used in training | ROC-AUC | TPR at 1% FPR | TPR at 10% FPR |
|---|---|---|---|---|
| 250 | 100% | 0.850 (0.496) | 0.291 (0.037) | 0.622 (0.121) |
| 500 | 100% | 0.918 (0.495) | 0.395 (0.031) | 0.773 (0.131) |
| 1,000 | 100% | 0.909 (0.418) | 0.335 (0.001) | 0.748 (0.058) |
| 1,100 | 91% | 0.907 (0.428) | 0.279 (0.006) | 0.746 (0.060) |
| 1,200 | 83% | 0.890 (0.425) | 0.230 (0.005) | 0.686 (0.057) |
| 1,500 | 67% | 0.847 (0.420) | 0.148 (0.005) | 0.583 (0.054) |
| 2,000 | 50% | 0.793 (0.380) | 0.047 (0.005) | 0.341 (0.031) |
| 4,000 | 25% | 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
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.”
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.
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.
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.
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.
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)
External audit (Isolation Forest): ROC-AUC, target (blind)