acceptodds
Under review as a conference paper at ICLR 2027

Amplifying Privacy Attacks on Tabular Foundation Models via a Stolen Surrogate

Abstract

Tabular foundation models (TFMs) achieve state-of-the-art performance on structured data. The fact that these models are usually adapted for their sensitive downstream tasks by in-context learning raises the question of privacy risks. We show that even a query-only adversary can extract a significant amount of private information from the context examples by first stealing a surrogate of the preconditioned model and then running privacy attacks on the copy. Without a single sample from the in-context data distribution, without knowledge of the context size, and without any write access to the context, the adversary sends random queries, adapts a public TFM checkpoint to the answers, and obtains a high-fidelity surrogate. The adversary can then condition on arbitrary contexts and query the surrogate without any restrictions. We show that the surrogate can significantly amplify the risks of standard privacy attacks, including membership inference, attribute inference, and property inference. We analyze our method across various settings, including different datasets, TFM backbones, context sizes, and output formats. We find that the combination of stealing and privacy attacks generalizes over these settings, and that its success grows with task difficulty. Restricting the output to the predicted label is still vulnerable to theft, but prevents privacy attacks on a classification task, whereas private information from regression still leaks. Our anonymous codebase is available at https://anonymous.4open.science/r/tfm-priv-attack-0BFC.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
tabular foundation model, in-context learning, membership inference, attribute inference, property inference

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

The paper investigates privacy risks in tabular foundation models (TFMs) by adapting model-stealing techniques to models conditioned on private in-context datasets. The proposed approach allows an attacker with black-box query access to construct surrogate models without requiring auxiliary data from the victim's distribution. Specifically, the attacker generates random queries based on the public data schema and trains surrogate models to reproduce the victim's predictions. The authors then propose membership inference (MIA), attribute inference (AIA), and property inference (PIA) attacks that leverage the stolen surrogates to extract information about the victim's private in-context dataset. Experiments across several tabular classification and regression datasets demonstrate that model stealing poses a privacy risk to TFMs and can improve the effectiveness of privacy attacks compared to the considered non-stealing baselines.

Critical strengths
  1. The paper addresses an interesting and relevant privacy problem and is the first to investigate model-stealing attacks against tabular foundation models (TFMs). The idea of exploiting the victim's predictive capabilities to generate pseudo-labeled data for training surrogate models is clever, as it removes the restrictive assumption that the attacker has access to an auxiliary dataset drawn from the victim's data distribution.
Critical weaknesses
  1. The paper lacks sufficient algorithmic detail, making several components of the proposed methodology difficult to understand and reproduce. None of the attacks is presented through a sufficiently precise algorithmic description, either in the main text or in the appendix. For instance, the model-stealing procedure is described as minimizing the KL divergence between the victim's and surrogate's predictions, with an additional discussion of optimizing the surrogate's context in the embedding space. However, the exact optimization procedure and its implementation are not clearly specified. Similar issues arise in the descriptions of the privacy attacks and their additional components, such as the neighborhood-based refinement procedure used for MIA. These algorithmic aspects should be formalized more carefully to ensure reproducibility.

  2. The neighborhood-based refinement procedure used in MIA requires more than 32,000 additional queries to the victim model for each target record, while providing only a marginal improvement in attack performance. This raises concerns about the practicality and scalability of the attack, particularly when targeting a large number of records or interacting with a rate-limited API. Similarly, the model-stealing stage requires a large number of queries to achieve high fidelity.

  3. The proposed attacks are largely motivated by intuition, without sufficient theoretical or empirical evidence supporting the assumptions underlying their effectiveness. For MIA, the surrogate models are trained to reproduce the predictions of a victim that may already contain the target record in its context. Consequently, information about the target record may already be encoded in the surrogate, even under the non-membership hypothesis, raising questions about the validity of the constructed IN and OUT distributions. For AIA, it is unclear why the victim should systematically exhibit higher relative confidence than the surrogate on the correct attribute completion, especially given that the surrogate is explicitly trained to approximate the victim's predictive distribution. A similar concern applies to PIA, where the attack assumes that predictions for which the victim is more confident than the surrogates are more informative about the private context. The authors do not sufficiently justify this assumption or the choice of retaining only the top 10% of ranked samples. Additional theoretical motivation or empirical analyses validating these assumptions would considerably strengthen the proposed methodology.

Questions for authors
  1. The MIA constructs in and out distributions based on the presence or absence of the target record in the surrogate's context. However, if the target is a member, the surrogates may have already learned information about it during model stealing. How could this affect the reliability of the attack?
  2. Could the authors clarify how the decision threshold τ\tau is selected in the MIA? In particular, how can the attacker calibrate this threshold without access to known member and non-member records, as assumed by the threat model?
  3. Could the authors provide further intuition or empirical evidence explaining why the victim should exhibit higher relative confidence than the surrogates on the correct attribute completion?
  4. Could the authors provide a clearer explanation of the rationale behind the rank score and the selection of the top 10% of probes in the property inference attack?
  5. Could the authors further explain the results in Figure 4? Given that the surrogates are trained to approximate the victim's behavior, why does an architectural mismatch sometimes improve attack performance? For example, PFN surrogates consistently achieve slightly higher AIA accuracy, while MIA against PFN victims performs better with mismatched surrogates. PIA also occasionally benefits from architectural mismatch. What could explain these differences?
  6. Could the authors provide more details on the types and cardinalities of the attributes targeted by AIA, including the treatment of continuous attributes? This is important for interpreting reconstruction accuracy, since an accuracy of 30% could be below the random-guessing baseline for a binary attribute.
4 comments

We thank the reviewer for recognizing that our paper "is the first to investigate model-stealing attacks against tabular foundation models" and for finding our pseudo-labeling idea "clever", as it removes the need for an auxiliary dataset from the victim's distribution. We address their comments and questions below.

W1. The attacks lack precise algorithmic descriptions.

We added pseudocode for all components in App. B.

W2. Is the attack practical against a rate-limited API?

To address the reviewer’s comment, we made the unrefined attack our default. After stealing, it needs one victim query per target, and refinement is now an option (App. C.2). We also ran the attack at smaller stealing budgets B (Fig. 5) and counted all victim queries, sharing the stealing cost over the 10,256 candidates of a membership pool (Table 7):

AttackQueries per candidateTPR@5%FPRAUC
Non-stealing baseline25723.2%.712
Ours, B = 1,0241.125.0%.724
Ours, B = 16,3842.631.5%.779
Ours, B = 65,536 (default)7.432.7%.781
Ours + refinement (optional)32,77539.4%.812

Without refinement, our default attack uses 35 times fewer queries per candidate than the non-stealing baseline and still reaches a higher TPR@5%FPR. Stealing also works with fewer queries: fidelity is already .825 at B=210B = 2^{10} and .934 at B=214B = 2^{14} (Fig. 2(b)), and at these budgets all three privacy attacks still match or outperform their non-stealing baselines (Fig. 5). Our default budget fits in 128 requests to SAP's RPT API, below its limit of 1,000 requests per hour (SAP).

W3. Is there evidence for the assumptions behind the attacks?

To address the reviewer’s comment, we tested the assumption behind each attack and added the results to App. C.7 (Table 12). All three hold: the surrogates absorb only about 15% of the membership signal, the victim's lead over the surrogates is larger at the true completion than at the wrong ones in every run, and the probes that RANK keeps lie closer to the private context. We give the details of each attack below.

W3 (MIA) / Q1. Do the surrogates already encode the target under OUT?

We agree that the surrogates pick up some information about the private context during stealing, as is known for distillation (Jagielski et al., 2023). In theory, this weakens our attack, since the OUT surrogates then carry a slight membership signal themselves. In practice, a record in the victim’s context has a much stronger effect. The victim is 1.27 logits more confident on members than on non-members, the surrogates only 0.19, so they reproduce about 15% of the membership signal (App. C.7). Our attack therefore still reaches 32.7% TPR@5%FPR, against 23.2% for the non-stealing baseline.

W3 (AIA) / Q3. Why is the victim more confident than the surrogates on the true completion?

The surrogates reproduce the victim’s overall function, but they are fitted on queries far from the real records (App. C.6), so they miss the effect of an individual context record. In every run, the gap between the victim’s and the surrogates’ confidence is larger at the true completion than at the wrong ones. As a result, our attack ranks the true completion first for 68.4% of the targets, against 14.7% for a random guess.

W3 (PIA) / Q4. Why rank probes by victim–surrogate confidence, and why keep the top 10%?

Most schema-sampled probes lie far from the private context, where the victim behaves like one more surrogate. RANK keeps the probes where the victim is more confident than most surrogates, and these lie closer to a context record in the victim’s representation (Table 12). The 10% is not critical: every kept fraction from 90% down to 10% improves on the non-stealing baseline (App. C.5, Table 8).

Q2. How is the MIA threshold set without known members?

The attack needs no labeled records: the adversary scores the unlabeled candidate records it wants to test and flags the top α\alpha of them as members. We use membership labels only to evaluate this rule: with α=5\alpha = 5\\%, it reaches 28.3% TPR at 3.2% FPR, against 17.5% TPR at 3.4% FPR for the non-stealing baseline. TPR@5%FPR and AUC in the paper summarize the attack over all thresholds, the standard way to report membership inference (Carlini et al., 2022).

Q5. Why does a mismatched surrogate sometimes do better (Fig. 4)?

We reran Fig. 4 with three seeds, a larger membership pool and six backbones. With these changes, a mismatched surrogate no longer helps or hurts the attack in a consistent way. It beats the matched surrogate in 17 of the 30 pairs, which is close to chance. We hypothesize that this is because a surrogate only has to reproduce the victim's predictions on the transfer set, and several backbones can do that equally well. The amount of leakage depends mainly on the victim, since some backbones react much more strongly to a single context record than others. Table 13 (App. C.7) shows that, for a fixed victim, the membership signal varies by at most 0.17 logits across surrogates, whereas across victims it ranges from 0.64 to 2.57 logits. TPR@5%FPR behaves the same way. It varies by at most 14.7 points across surrogates but ranges from 10.7% to 77.4% across victims.

Q6. Which attributes does AIA target, and what is chance accuracy?

To address the reviewer’s question, we added a random-guessing baseline (Rand) to Table 2(a). The unknown attributes are drawn at random from all feature columns, and the label is always part of the known anchor. Across the 45 datasets, the attacked attributes are:

Attribute typeShareCandidatesRandom guess
Binary10.5%2 categories1/2
Categorical23.2%3 to 7,518 categories (median 5)1/k
Numeric66.3%16 bins over the declared range1/16

At ∣U∣=1\vert U \vert = 1, our attack recovers 67.7% of the unknown attributes, against 15.4% for Rand. On binary attributes alone, it reaches 89.5%, against 50% for random guessing.

If our answers have addressed the reviewer’s questions, we would greatly appreciate if they could update their score.

As the rebuttal period is coming to an end, we would like to thank the Reviewer once again for their constructive feedback. To address their comments, we have additionally:

  1. Added pseudocode for model stealing and all three privacy attacks (App. B).
  2. Made per-record refinement optional, so that our attack requires only 7.4 victim queries per candidate, including stealing, i.e., 35x fewer than the non-stealing baseline, and still attains a higher TPR@5%FPR (32.7% vs. 23.2%) (App. C.2, Table 7).
  3. Validated the assumption underlying each attack: the surrogates reproduce only 15% of the victim's membership signal, the victim's confidence lead is larger at the true AIA completion in every run, and the probes selected by RANK lie closer to the private context (App. C.7, Table 12); the 10% threshold is not critical, as every kept fraction from 10% to 90% improves on the baseline (App. C.5).
  4. Extended the backbone-mismatch study to three seeds and six backbones, where a mismatched surrogate outperforms the matched one in 17 of 30 pairs, i.e., at chance level (App. C.7, Table 13).
  5. Added a random-guessing baseline for AIA, which our attack far exceeds (67.7% vs. 15.4%; 89.5% vs. 50% on binary attributes) (Table 2(a)).
  6. Clarified that MIA requires no labeled records: flagging the top 5% of candidates as members yields 28.3% TPR at 3.2% FPR, compared with 17.5% at 3.4% for the baseline.

We believe these additions improve the reproducibility of our attacks and provide empirical support for each attack's assumptions. We would appreciate your reconsideration of the rating, including an increase if the new experiments address your concerns.

I thank the authors for their detailed rebuttal and the substantial improvements to the manuscript.

The added pseudocode improves reproducibility. The additional empirical analyses provide enough evidence. The experiments on architectural mismatch and the random-guessing baseline for AIA also satisfactorily address my questions.

I still have a minor reservation regarding MIA calibration, particularly whether the constructed IN and OUT distributions accurately reflect the membership hypotheses. Nevertheless, the additional empirical evidence helps alleviate this concern.

Overall, I appreciate the authors' efforts and consider the rebuttal substantial and convincing on most points. The additional analyses and improvements have strengthened my confidence in the proposed approach. I am therefore inclined to increase my score.

Reviewer_wQ4iOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the privacy leakage for tabular foundation models' deployment. Specifically, the private information in the TFM's context may be extracted by attackers, even though the attackers can only query the TFM in a black-box way. The paper shows that by randomly generating many queries and obtaining the probability vectors of the victim model, the attacker can construct the transfer set. Then, the attacker uses its local surrogate model to optimize both the model weights and the context to close the distribution gap between the surrogate and the victim model. And then perform the privacy attacks once the victim functionality is stolen. The authors demonstrate the effectiveness of the attack on two TFMs and one dataset benchmark.

Critical strengths

This work studies a new stealing attack that can steal the functionalities of a victim TFM using black-box queries. So conceptually, I think this paper reveals a threat and an attack opportunity that can inspire more secure deployment of TFM in practice.

Critical weaknesses

I have the following concerns on assumptions and contributions that lead to a negative overall assessment.

  • Technically, I think this paper's contribution is limited. The techniques used are straightforward to me in this specific setup, i.e., in this surrogate setting, the default attack would be first approximate the victim model and then perform the privacy attack. So I think the conceptual novelty would be revealing such attack possibilities in this setting.
  • I have following concerns on this settings' assumptions:
  1. Is it reasonable to assume that the attacker can obtain a TFM whose size and capability is comparable to the victim? What if the victim model is much larger, and the surrogate is much smaller?
  2. The context size in the experiments is very small, if you use a much larger context, can you still perform the attacks successfully?
  3. When constructing the transfer set, the attacker needs to submit a large amount of "random" queries to the victim model. Isn't this behavior very easy to flag? As the large amount of random queries may be easily distinguished from normal behaviors.
Questions for authors

Can the authors 1) clarify the technical contributions and 2) answer my questions above on the rational of the attack setting and (implicit) assumptions?

2 comments

We thank the reviewer for recognizing that our work "reveals a threat and an attack opportunity that can inspire more secure deployment of TFM in practice" and for finding the paper "clear and easy to follow". We address their comments and questions below.

W1 / Q1. What are the technical contributions?

While stealing and then attacking is the natural approach and its building blocks exist, we are the first to combine them into a practical, effective attack against in-context TFMs: prior black-box attacks performed near random guessing (Wahdany et al., 2026), because shadow models need in-distribution data. Our stolen surrogates remove this assumption and still find significant leakage (32.7% TPR@5%FPR). Two design choices make this possible:

  1. Stealing without real data. We query the victim on records sampled from the public schema and fit the surrogate's context tokens together with its weights (.952 fidelity, vs. .841 unfitted).
  2. Surrogates as shadow models. We get the IN hypothesis by appending the target to a surrogate's context, which is impossible on the victim. Without fitting, MIA falls below the non-stealing baseline (17.1% vs. 23.2% TPR@5%FPR).

W2a. What if the surrogate is much smaller than the victim?

Our threat model does not assume that the adversary knows the victim's backbone. We extended the mismatch study (Fig. 4) to all pairs of six backbones, from the largest, TabFM (1.6B parameters), to the lightest, Causilo (36M). Against a TabFM victim, a Causilo surrogate still reaches a membership TPR@5%FPR of 29%. Overall, the surrogate's backbone changes the attacks only slightly, and the surrogate that matches the victim is best for only one of the six victims. The victim's backbone matters far more: membership TPR@5%FPR ranges from 11% to 77% across victims.

W2b. Do the attacks work with a much larger context?

To address the reviewer’s comment, we extended the context-size study in Fig. 4(e)-(h) to m,n∈64,128,256,512,1,024m, n \in \\{64, 128, 256, 512, 1{,}024\\} for both the victim's context and the surrogate's prompt. Leakage decreases as the victim's context grows, as expected because each record has less influence on the prediction, but at m=1,024m = 1{,}024 membership inference still reaches a TPR@5%FPR of 22.0%, against 9.0% for the non-stealing baseline. A wrong prompt size nn changes fidelity by at most 0.006 and lowers membership TPR@5%FPR by only 0.7 points on average. In addition, we ran the attack at m=n=4,096m = n = 4{,}096 and found leakage similar to that at m=n=1,024m = n = 1{,}024: fidelity is .875 (vs. .845), membership inference reaches a TPR@5%FPR of 20.6% (vs. 22.0%).

W2c. Aren't the many random queries easy to flag?

  1. The volume is within normal use of deployed tabular APIs. Tabular prediction APIs are built for batch scoring. For example, SAP's public API for SAP-RPT-1 accepts up to 512 query rows per request and 1,000 requests per hour per user (SAP documentation). Our default budget of 2162^{16} rows therefore fits in 128 requests, about 13% of a single user's hourly quota. Query volume alone therefore does not separate the adversary from a user scoring a large table.

  2. Far fewer queries still enable an effective attack. Fig. 2(b) shows that the surrogate reaches .825 fidelity with only 1,024 queries (2 requests to the API above) and .934 with 16,384 (32 requests), against .952 at our default 2162^{16}. At B=214B = 2^{14}, membership inference reaches 31.5% TPR@5%FPR, close to the 32.7% at 2162^{16} and above the 23.2% of the non-stealing baseline. At B=210B = 2^{10}, it is on par with the baseline (25.0%).

If these answers have addressed the reviewer’s comments, we would greatly appreciate if they could update their score.

As the rebuttal period is coming to an end, we would like to thank the Reviewer once again for their constructive feedback. To address their comments, we have additionally:

  1. Clarified our contribution: prior black-box attacks on TFM contexts perform near random guessing because shadow models require in-distribution data (Wahdany et al., 2026), whereas our fitted surrogates require no such data and reach 32.7% TPR@5%FPR, compared with 17.1% without fitting (App. C.1, Table 6).
  2. Evaluated all pairs of six backbones from 36M to 1.6B parameters, where even a 36M Causilo surrogate reaches 29% TPR@5%FPR against a 1.6B TabFM victim, and leakage varies far more across victims than across surrogates (Fig. 4).
  3. Scaled the private context to 1,024 rows, where MIA still reaches 22.0% vs. 9.0% TPR@5%FPR for the non-stealing baseline (Fig. 4(e)–(h)); an additional run at 4,096 rows, reported in our response, yields 20.6%.
  4. Showed that the query volume is consistent with ordinary API usage; our default budget fits in 128 requests, about 13% of a single user's hourly quota on SAP's RPT API, and 32 requests (B=214B = 2^{14}) already yield 31.5% TPR@5%FPR (Fig. 2(b), Fig. 5).

We believe these additions resolve the questions on surrogate size, context size and the detectability of the attack. We would appreciate your reconsideration of the rating, including an increase if the new experiments address your concerns.

Reviewer_1cceOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies attacks on the private context of a tabular foundation model (TFM). The attacker knows the column types, categories, and numeric ranges. It cannot read or change the victim's context and has no real auxiliary samples from the same distribution. It collects API responses and trains local surrogate models by updating their context embeddings and weights. These models support membership inference (MIA), attribute inference (AIA), and property inference (PIA). Evaluation on 45 TabArena datasets shows high predictive fidelity and stronger attacks than direct baselines. An important observation is that a surrogate can preserve useful privacy signals despite the attacker's limited access to the original context.

Critical strengths
  1. A useful empirical extension to private TFM contexts. The paper demonstrates attacks without auxiliary in-distribution records or write access to the victim context. This supports a conditional privacy warning.

  2. Reasonably broad evaluation and an informative refinement ablation. Evaluation spans 45 classification/regression datasets and several attack types. Table 5 (p. 17) shows that unrefined MIA retains most of the refined attack's benefit: TPR@5%FPR is .340 versus .359, with the direct baseline at .217.

Critical weaknesses
  1. The technical novelty is limited. The method uses model extraction and LiRA-style shadow-model scoring, but does not show whether context optimization improves MIA over weight-only distillation. Also, the schema categories and ranges come from the full benchmark table, and the paper does not measure the distribution gap between query data and real data. The paper does not explain why context optimization helps MIA or how the query-data distribution affects the attack.

  2. The attack requires a large query budget for a limited gain. It first spends B=216=65,536B=2^{16}=65,536 queries to train the surrogates, then one query per candidate. For 256 candidates, the nominal total is 65,792 queries to identify about 44 of 128 members at 5% FPR. At 5% FPR, the non-stealing baseline has a TPR of 21.7%, compared with 34.0% for unrefined MIA and 35.9% for refined MIA (Tables 1 and 5). Thus, unrefined MIA improves TPR by 12.3 percentage points, and refinement adds another 1.9 points. The paper shows higher attack success, but does not show whether this advantage remains when compared to other attack baselines under the same total query budget, including the queries used to train the surrogates.

Questions for authors
  1. What does context optimization add? At the same query budget, what are the MIA TPR@5%FPR and AUC when training only weights, only context tokens, or both? Figure 2(a) reports .952 fidelity for weight-only training and .949 for joint training, but gives no MIA comparison. Please use the same IN/OUT scoring for these tests.

  2. What is the total cost of each attack? Please compare MIA TPR/AUC at the same total number of distinct victim-query rows. Count extraction and all neighbor or perturbation queries. Why does Table 5 count only one query for the direct baseline that uses Hamming-1 neighbors? How well does unrefined MIA work at B=1,024 or 16,384? Please also give the training time for 32 copies and the number of candidates that share the extraction cost.

  3. Please clarify the recovery claims and service setup. Which PIA results use regions selected from the private context? Does MIA still improve on the baseline with eight TabICL preprocessing views rather than one? Please describe a service where the private context stays fixed while the attacker sends these synthetic rows.

  4. Does the attack work against a DP-protected TFM? Table 3 tests label-only and top-1-label-plus-confidence classification responses, and point-estimate regression responses. It does not test differential privacy. If a method such as the cited TabPATE fits this setup, please report prediction utility, attack TPR/AUC, total queries, and the privacy budget used across repeated responses. Otherwise, please state that DP-protected TFMs were not tested.

  5. How close are the query data to the real data? Please compare the schema-uniform queries and the response-guided perturbations with real rows separately. Report each feature's distribution, how well a two-sample classifier can tell the query and real rows apart, and exact/near-duplicate rates. Which categories and ranges taken from the full table would actually be public? Does the attack still work when these are set independently of the benchmark data?

6 comments

We thank the reviewer for recognizing our work as a “useful empirical extension to private TFM contexts” and our broad evaluation. We address their comments and questions below.

W1. The technical novelty is limited.

While model extraction and shadow-model scoring are established techniques, we are the first to combine them into a practical, effective attack against in-context TFMs: prior black-box attacks performed near random guessing (Wahdany et al., 2026), because shadow models need in-distribution data. Our stolen surrogates remove this assumption and still find significant leakage (32.7% TPR@5%FPR, where random guessing gives 5%). Two design choices make this possible:

  1. Stealing without real data. We query the victim on records sampled from the public schema and fit the surrogate's context tokens together with its weights (.952 fidelity, vs. .841 unfitted). Q1 below shows what the context tokens add for MIA.
  2. Surrogates as shadow models. We get the IN hypothesis by appending the target to a surrogate's context, which is impossible on the victim. Without fitting, MIA falls below the non-stealing baseline (17.1% vs. 23.2% TPR@5%FPR).

Q1. What does context optimization add?

To answer this, we ran the ablation the reviewer asked for. All variants use the same B=216B = 2^{16} queries and the same IN/OUT score (Eq. 1). The only difference is which parts of the surrogate we fit. We added the results to App. C.1 (Table 6):

Fitted partFidelityMIA TPR@5%FPRMIA AUC
Non-stealing baseline–23.2%.712
Weights only (θ~\tilde\theta).95131.2%.692
Context tokens only (tt).92731.5%.762
Both (t+θ~t + \tilde\theta, ours).95232.7%.781

With weights only, the surrogates match the victim functionality just as closely as with both (fidelity .951 vs. .952), but MIA AUC drops to .692, below the non-stealing baseline (.712). Fitting the context tokens as well brings AUC up to .781 and TPR@5%FPR from 31.2% to 32.7%. Fitting both keeps the fidelity of weight-only fitting and gives the highest MIA AUC and TPR of all variants. We also report results for attribute and property inference in App. C.1.

W1 (cont.) / Q5. How close are the queries to the real data?

Distance between queries and real data. To answer this, we compared our queries with the real records and added the analysis to App. C.6. Half of the queries are sampled uniformly from the schema. The other half, the response-guided queries, are perturbations of the uniform queries to which the victim gives the highest probability for rare classes. For each half, we measure how far each attribute's distribution is from the real records (total variation distance for categorical attributes, Wasserstein distance divided by the attribute's range for numeric ones) and average over the attributes. A classifier tests whether queries and real records can be told apart. For scale, we compare two disjoint samples of real records in the same way:

Compared with real recordsCategorical (TV)Numeric (Wasserstein)Classifier acc.
Other real records.022.005.501
Uniform queries.304.284.998
Response-guided queries.341.251.998

Both halves are 14 to 57 times farther from the real data than two real samples are from each other, and the classifier separates them from real records almost perfectly. Apart from one dataset whose entire domain is smaller than our query budget, no query equals a real record, and in 91% of runs no private context record has even a Hamming-1 neighbor among the queries.

We also compared queries and real records in the victim TFM's internal representation in Table 11. There, too, each private context record is about 10 times farther from its nearest query than from its nearest real record, and a linear classifier separates queries from real records with an accuracy of at least .997.

Which schema information is public. Model-serving frameworks let the provider declare the valid categories and numeric range of each input, publish them in the API documentation, and reject requests outside them (Porter, BentoML). Hence, for our threat model, we assume that the adversary knows the schema. In our experiments, each numeric range is the attribute's minimum and maximum over the full dataset.

To test whether the attack needs these exact ranges, we attack a TabICLv2 victim on students_dropout with every range widened to 1.25 and 1.5 times this span (App. C.8, Table 14). We also let the adversary recover the ranges from the victim alone, by finding where it rejects inputs or where its prediction stops changing. We keep the categories fixed, since probing recovers them exactly:

Numeric rangesFidelityMIA TPR@5%FPR
Full-dataset min/max (default).90239.1%
Widened to 1.25–1.5× the span.886–.89538.2–38.5%
Recovered by probing.899–.91238.0–39.4%

Fidelity changes by at most .016 and TPR@5%FPR by at most 1.1 points, so the attack needs neither an exact schema nor queries close to the private data.

W2 / Q2. What does each attack cost, and how does MIA do at equal or lower budgets?

To address the reviewer’s question, we ran additional experiments where we varied the stealing budget B and counted every victim query of each attack (Table 7, Fig. 5). We also made the unrefined attack our default, with refinement as an option (App. C.2). The table divides the total queries by the 10,256 candidates of a membership pool:

AttackQueries per candidateTPR@5%FPRAUC
Non-stealing baseline25723.2%.712
Init-full (no fitting), B = 65,5367.410.4%.611
Ours, B = 1,0241.125.0%.724
Ours, B = 16,3842.631.5%.779
Ours, B = 65,536 (default)7.432.7%.781
Ours + refinement32,77539.4%.812

Our default attack uses 35 times fewer queries per candidate than the baseline and reaches a higher TPR@5%FPR (32.7% vs. 23.2%). At the same total budget, Init-full, which conditions the public model on all queried records without fitting it, reaches only 10.4%, so the gain comes from fitting the copies, not from the queries alone. Stealing also works with fewer queries: fidelity is already .825 at B=210B = 2^{10} and .934 at B=214B = 2^{14} (Fig. 2(b)), and at these budgets all three privacy attacks still match or outperform their non-stealing baselines (Fig. 5).

W2 / Q2 (cont.). Why does Table 5 count one query for the baseline?

Thank you for catching the error in Table 5. The baseline also queries the target's 256 Hamming-1 neighbors, so it costs 257 queries per target, not one. We corrected the count (now Table 7).

W2 / Q2 (cont.). How long does fitting take, and how many targets share its cost?

The stealing queries and the 32 copies are reused for all 10,256 membership candidates, the 100 attribute-inference targets and every property-inference region of a dataset. Fitting the 32 copies takes about 2.4 GPU-hours and needs no victim queries. The default budget fits in 128 requests to SAP's RPT API, below its limit of 1,000 requests per hour (SAP).

Q3. Which PIA results use regions selected from the private context?

The regions with ℓ≥1\ell \geq 1 conditions are selected from the private context. The ℓ=0\ell = 0 results, which estimate the label proportion of the whole context, use no region selection.

The region is part of the challenge given to the adversary, in the same way that membership inference hands the adversary a real record and asks whether it is in the context. The adversary is not asked to find a region, but to estimate the label proportion α\alpha within a given one, e.g., a subgroup they are interested in. Each region is a conjunction of ℓ\ell conditions of the form "attribute = category". We select the conjunctions that occur most frequently in the context. The non-stealing baseline and the oracle are evaluated on exactly the same regions, so the comparison between methods does not depend on this choice.

We made this explicit in Sec. 4.1 (property inference) and App. A.

Q3 (cont.). Does MIA still beat the baseline with eight TabICL views?

To address the reviewer’s question, we ran an additional experiment, where we use TabICLv2's default ensemble of 8 members. We added the results to Sec. 4.2 (Table 4).

Victim membersVictim acc.AttackFidelityMIA TPR@5%FPRMIA AUCAIA acc.PIA MAE
1 (default).745Baseline.58610.8%.640.290.219
1 (default).745Ours.90239.1%.735.618.204
8.749Baseline.48510.7%.642.280.232
8.749Ours.90137.7%.742.632.218

Our results show that the MIA still improves significantly over the baseline with 8 TabICL preprocessing views.

Q3 (cont.). Which service keeps the context fixed while accepting synthetic queries?

Two deployed services work this way, and we added both to Sec. 1. In SAP's tabular orchestration service for SAP-RPT (SAP Community), the provider registers its table once, and later prediction requests contain only the query rows. Prior Labs' TabPFN API does the same: the fit is stored on the server, and later calls send only the query rows and a model identifier (PriorLabs/tabpfn-client). Our adversary simply sends schema-valid rows as prediction requests, like any other user.

Q4. Does the attack work against a DP-protected TFM?

Since DP offers post-processing guarantees, repeated responses do not incur additional privacy costs, and the privacy budget is spent once during training.

To address the reviewer's comment, we ran our attacks (functionality stealing, membership, attribute and property inference) against a TabPATE-protected TabICLv2 victim at ε=1\varepsilon = 1 (δ=10−5\delta = 10^{-5}) on the credit-card-default dataset and compared it with the undefended victim (App. C.12, Table 15):

VictimTest AUCStealing agreementMIA TPR@5%FPRMIA AUCAIA acc. (∣U∣=1\vert U \vert = 1)PIA MAE (ℓ=0\ell = 0)ε\varepsilon (δ=10−5\delta = 10^{-5})
No DP.77697.7%32.8%.713.58.213∞\infty
TabPATE.75495.8%6.8%.498.31.2801.0

TabPATE stops membership inference (AUC .498), brings attribute inference down to what the victim alone gives (.31 vs. .30) and makes property inference less accurate, at a small cost in utility. DP does not prevent functionality stealing (95.8% agreement), since it protects individual records, not the model's function. The budget stays at ε=1\varepsilon = 1 over all 206k queries.

If our answers have addressed the reviewer’s questions, we would greatly appreciate if they could update their score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the rebuttal

We thank all reviewers for their constructive feedback. We are grateful that Reviewer iWFB recognizes our paper as "the first to investigate model-stealing attacks against tabular foundation models" and finds our pseudo-labeling approach, which removes the need for auxiliary data from the victim's distribution, "clever". We thank Reviewer 1cce for considering our work "a useful empirical extension to private TFM contexts", and Reviewer wQ4i for stating that it "reveals a threat and an attack opportunity that can inspire more secure deployment of TFM in practice".

Additional Highlights of Our Rebuttal

To incorporate the reviewers' feedback, we conducted a large set of additional experiments and summarize the highlights below:

  1. Lower query budgets: We made the per-record refinement optional, so after stealing, our attack needs only one victim query per target. With B=214B = 2^{14} queries instead of 2162^{16}, membership inference reaches 31.5% TPR@5%FPR, and with B=210B = 2^{10} it is on par with the non-stealing baseline.
  2. Fitted-parts ablation for the privacy attacks: We showed that fitting the surrogate is necessary. Without it, membership inference falls below the non-stealing baseline (17.1% vs. 23.2% TPR@5%FPR), while fitting both the context and the weights raises it to 32.7%, validating the design choice of our attack.
  3. Larger mismatch in backbone and context size: We ablated the impact of different backbones and context sizes on the attack success. Across six backbones from 36M to 1.6B parameters and context sizes from 64 to 1,024 rows, knowing the victim's backbone or context size gives the adversary no advantage, and all attacks still leak even at the largest context.
  4. Defenses and deployment variants: We evaluated our attack against target models protected with differential privacy, using TabPATE at ε=1\varepsilon = 1. TabPATE reduces membership inference to chance, while functionality stealing still succeeds. Serving the victim as an ensemble, or declaring wider numeric ranges than the actual data span, barely changes the attack performance.
  5. The attack does not rely on real data: We build our queries from the public schema alone, and a classifier can tell them apart from real records almost perfectly. Since the surrogates never see anything close to the private records, they copy the victim's function but not its membership signal, and our attacks exploit this gap.

We believe that these additions have substantially strengthened our submission, and we remain available should further questions arise.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.