acceptodds
Withdrawn from ICLR 2027

From Row-Level to Table-Level: Zero-Shot Private Attribute Inference in Tabular Data with Large Language Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Withholding sensitive columns is a common precaution when sharing tabular data, yet the remaining attributes may still support inference of the hidden values. We study this privacy risk under a zero-shot adversary that observes the released dataset and uses pretrained large language models (LLMs), without receiving ground-truth examples of the target attribute. The attack is evaluated under different levels of feature disclosure and with or without aggregate target proportions. Beyond direct LLM predictions for individual samples, our analysis asks whether statistical dependencies across the dataset can enable additional attribute recovery. We develop an evidence-based pseudo-context strategy that uses the LLM's scores to select pseudo-labeled samples for a tabular learner, deriving a separate context budget for each predicted class without ground-truth target labels. On demographic and health survey data, direct LLM inference is competitive with zero-shot baselines, and the table-level extension improves average accuracy and macro recall over direct prediction. These findings demonstrate how pretrained semantic knowledge can supply supervision for learning from an unlabeled dataset, exposing a source of privacy risk beyond independent sample-level inference.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Private Attribute Inference, Zero-Shot Prediction

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_QXPpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies a zero-shot adversary who observes a table with a sensitive column removed and attempts to infer its values without receiving ground-truth target labels. The main contribution is an adaptive rule for determining how many pseudo-labeled examples to select from each predicted class.

Critical strengths

The paper studies whether access to an entire released table enables more inference than inspecting its rows independently. It distinguishes individual feature disclosure from access to aggregate target proportions.

The evaluation includes two datasets, four ACS states, several sensitive attributes, multiple LLMs, and both classification and geographic prediction.

Critical weaknesses

This paper uses zero-shot LLM predictions as pseudo-labels to fit a downstream classifier, which is an established pattern in NLP. Examples include self-training on zero-shot classifier outputs (e.g., Gera et al., 2022) and LLM-as-annotator distillation. The related work frames pseudo-labelling only through Noisy Student and FixMatch.

The attacker is assumed to has no labelled examples, but the full ACS and NHIS public files, with all these attributes included, are available. An attacker would simply train a supervised model on that data and do much better than either E0 or E1. The paper also does not discuss whether predicting someone’s attributes from general population patterns should count as a privacy harm to the people in the released table.

E1 learns patterns from unlabeled rows, but any unlabeled samples from the same population would provide the same patterns. To show that releasing the table adds risk, the authors could build the context from a different group of people and see whether the gain remains. They could also test whether the gain grows as the released table becomes larger.

All the results come from deterministic selection with seed 42 on test sets of 1,200 samples. The reported result differences are quite small and close to be noise.

Questions for authors

What are the key differences of this paper compared to the closely related methods?

How much of E1’s gain can be matched by simply rescaling the LLM scores, without training any model?

Could you provide per-class precision and recall for selected targets?

Does E1’s gain persist when the pseudo-context comes from a different, non-released cohort? How does it scale with cohort size?

Why does E1 help Qwen3-14B but mostly not GPT-4.1?

Can you explain the mismatch between Tables 9 and 19, and confirm which TabPFN version was used?

6 comments
AuthorsOpenReview ↗

Prior work and privacy setting (W1-W2, Q1)

We would like to express our sincere gratitude for your careful review and valuable comments and suggestions. Our point-by-point responses are given below.

Weakness 1 / Question 1.

This paper uses zero-shot LLM predictions as pseudo-labels to fit a downstream classifier, which is an established pattern in NLP. Examples include self-training on zero-shot classifier outputs (e.g., Gera et al., 2022) and LLM-as-annotator distillation. The related work frames pseudo-labelling only through Noisy Student and FixMatch.

What are the key differences of this paper compared to the closely related methods?

Response: We agree that zero-shot pseudo-labeling and teacher-to-student transfer are established ideas. We have added Gera et al. (2022) to Section 2.2 and do not claim that this general pipeline is new. The key differences are:

  1. Privacy question and information access. Gera et al. study zero-shot text classification. We study recovery of a private column hidden throughout a tabular dataset, comparing controlled feature disclosure and released target proportions. The question is what additional private information becomes recoverable under those access conditions.
  2. Fixed semantic predictor, separate statistical learner. Gera et al. use an NLI model to score text--label hypotheses, construct pseudo-labeled entailment/contradiction pairs, and fine-tune that model. We keep the LLM fixed and transfer its pseudo-labels to a supervised tabular learner operating on visible features. Final predictions combine that learner's outputs with the original LLM scores.
  3. Adaptive evidence budgets instead of fixed per-class selection. Gera et al. rank examples by the gap between their highest and second-highest scores and select a fixed number per predicted class. Our selector instead corrects for each class's background across the dataset and converts accumulated positive evidence into a separate context budget. E0 and E1 retain the same initial LLM outputs, isolating additional recovery through tabular transfer.

Our contribution is therefore the systematic tabular privacy study and this specific evidence-based transfer strategy, rather than pseudo-labeling itself.

Revised manuscript: Section 2.2; Section 1 (contribution positioning).

Weakness 2.

The attacker is assumed to have no labelled examples, but the full ACS and NHIS public files, with all these attributes included, are available. The paper also does not discuss whether predicting someone's attributes from general population patterns should count as a privacy harm to the people in the released table.

Response: We model a publisher who deliberately removes private columns before sharing a dataset. The attacker observes the retained fields but has no ground-truth examples of the withheld target. Complete ACS/NHIS files let us simulate this disclosure policy and evaluate recovery; their public availability is not a claim that the labels are inaccessible outside the experiment. When representative labeled data are available, supervised inference is the appropriate stronger-access reference.

Yes: population-based inference can compromise the intended confidentiality of a private attribute. Hiding a value does not protect it if the remaining features and general population knowledge still permit its recovery. Our study is a specific instance of the scenario you describe: attributes intentionally withheld by the publisher are inferred from disclosed features and pretrained knowledge. Such disclosure need not involve memorization of an individual's training data. We measure that attribute-recovery risk. This differs from attributing a formal, participation-specific privacy loss to publishing a particular individual; the external-cohort controls in W3/Q4 examine whether the statistical benefit requires those same released individuals.

Revised manuscript: Sections 3.1 and 6.

AuthorsOpenReview ↗

Score rescaling and model differences (Q2, Q5)

Question 2.

How much of E1's gain can be matched by simply rescaling the LLM scores, without training any model?

Response: We test whether a class-dependent adjustment of the existing scores can replace tabular learning. Let sics_{ic} be the original LLM score and bc=n−1∑isicb_c=n^{-1}\sum_i s_{ic} its mean across samples for class cc. On all 66 Qwen strict/weak tasks, we apply two fixed rules:

  1. Ratio rescaling: ric=sic/max⁡(bc,10−8)r_{ic}=s_{ic}/\max(b_c,10^{-8}). Dividing by the class mean reduces the relative advantage of classes that receive high scores across most samples.
  2. Positive-evidence rescaling: ric=[sic−bc]+/(1−bc)r_{ic}=[s_{ic}-b_c]_+/(1-b_c). Only support above the class mean is retained; scores at or below that mean contribute zero evidence. Degenerate columns are set to zero.

Each rule predicts arg⁡max⁡cric\arg\max_c r_{ic} for every sample; an all-zero evidence row retains E0. Neither rule constructs context or fits a model. E1 instead uses evidence to select pseudo-labels, learns feature dependencies, and fuses tabular predictions with the original scores. No target labels tune any rule. Mean Accuracy/Macro Recall changes from E0 (percentage points) are:

Dataset / conditionScore / class meanPositive evidenceE1
ACS strict−1.22 / +2.39+1.59 / +1.75+3.35 / +2.20
ACS weak−2.51 / +3.40+1.20 / +2.57+2.43 / +1.73
NHIS strict−8.47 / +0.04−3.13 / +0.09+0.22 / +0.71
NHIS weak−0.05 / +4.17+0.47 / +3.33+0.85 / +1.32

E1 achieves higher mean accuracy than both score-only transformations in all four groups, while also improving macro recall over E0 in each group. Thus the tabular learner adds recovery beyond these direct score adjustments. Both controls use across-sample class means, making this a comparison of score correction with learned transfer under the same information access. The existing permutation diagnostic in Table 21 further supports the role of feature–pseudo-label correspondence: retaining context rows and label counts while shuffling labels changes the ACS weak accuracy increment from +2.43 to −9.47 points.

Revised manuscript: Appendix, "Score-only controls" (transformations and matched results); Table 21, Appendix F.1 (existing permutation diagnostic).

Question 5.

Why does E1 help Qwen3-14B but mostly not GPT-4.1?

Response: GPT-4.1 provides stronger initial predictions and therefore leaves less recovery to the table-level step. In our 24 matched strict/weak tasks, 59.50% of the errors corrected by Qwen E1 were already correct under GPT E0. Nevertheless, E1 still recovers additional private attributes beyond GPT's own row-level predictions. This additional recovery is the relevant test of whether the table can provide useful information after a strong LLM has made its initial predictions.

Additional recovery with GPT-4.1. Across the seven ACS New York targets under strict disclosure, E1 raises GPT's mean accuracy from 47.32% to 48.24% (+0.92 percentage points). At the target level, INC improves from 73.17% to 79.08% (+5.92 points), and SCH from 38.58% to 40.83% (+2.25 points). On NHIS strict AGE, accuracy improves from 40.08% to 43.83% (+3.75 points). These are net improvements over GPT's own E0, after accounting for both corrected predictions and newly introduced errors. They use the same initial GPT outputs, no ground-truth target labels, and no additional LLM calls: the extra recovery comes from constructing pseudo-context and applying tabular inference.

Why some group averages decrease. E1 fits pseudo-labels as supervision. Their mistakes can therefore be learned as feature dependencies and propagated to samples whose E0 predictions were correct. Stronger E0 reduces correction opportunities but does not ensure that selection produces better training labels. For NHIS weak SCH, GPT context precision is 38.80%, versus 38.92% E0 accuracy, with only four of five classes represented: selection supplies noisy, incomplete supervision. The loss arises in tabular transfer, not simply fusion: GPT fused accuracy exceeds direct TabPFN accuracy in all four matched groups, while both remain below E0 in the negative groups (Appendix F.2, Tables 26-27). These comparisons are consistent with pseudo-label error propagation.

These results show that table-level inference can recover additional private attributes even after GPT-4.1's stronger initial predictions. Section 5.2 reports the matched comparisons, and Appendix F.2 analyzes the context quality and prediction changes associated with the different outcomes.

Revised manuscript: Section 5.2 (matched-model results); Appendix F.2, Tables 26-27 (context quality and task-level prediction changes).

AuthorsOpenReview ↗

Context source, dataset size, and stability (W3-W4, Q4)

Weakness 3.

E1 learns patterns from unlabeled rows, but any unlabeled samples from the same population would provide the same patterns. To show that releasing the table adds risk, the authors could build the context from a different group of people and see whether the gain remains. They could also test whether the gain grows as the released table becomes larger.

Response: We tested context from different people through all 168 cross-state transfers: four destination states, each using the other three states, seven targets, and strict/weak disclosure. Destination E0 scores and the 1,200 queried samples stay fixed. External-context mean accuracy/macro-recall increments are +3.04/+1.88 points under strict and +2.18/+1.39 under weak. On the identical samples outside the destination's original context, increments remain +3.50/+2.36 and +2.96/+1.56. The transfer therefore also works with pseudo-context from different people. The next control in Q4 addresses the more specific non-released-cohort question while holding population source fixed.

Effect of available table size. The size study fixes 200 queried samples and increases the available donor pool. Context is reselected at each size; donor count is not selected context size. Three fixed partitions across 66 tasks give 990 runs. Mean ACS accuracy increments over the same query E0 predictions are:

Donor samplesStrict ΔAccuracy (pp)Weak ΔAccuracy (pp)
100+1.15+1.39
200+1.61+1.86
400+2.48+1.90
800+2.92+2.16
1,000+3.18+2.18

From 100 to 1,000 donors, ACS macro-recall increments also rise from +0.00 to +1.85 points under strict and from +0.08 to +1.48 under weak. Increasing donor availability strengthens average ACS recovery on a fixed evaluation set.

Revised manuscript: Section 6; Appendix, Context source, available sample size, and backend seeds (cross-state and size controls).

Question 4.

Does E1's gain persist when the pseudo-context comes from a different, non-released cohort? How does it scale with cohort size?

Response: A non-released cohort means a separate auxiliary group available to the attacker but not included in the dataset being attacked. To test this distinction, we constructed a same-population, simulated-release comparison from three fixed partitions of the cached participants. Each partition defines 200 fixed queries and 400 internal donors; a disjoint 400-sample pool supplies external context. Queries never enter either context. Across 396 runs, ACS internal/external accuracy increments are +2.48/+2.85 points under strict and +1.90/+1.72 under weak. Both donor sources support additional recovery. This controlled partition changes release membership while retaining the population source and donor count.

Revised manuscript: Section 6; Appendix, Context source, available sample size, and backend seeds (disjoint donor controls).

Weakness 4.

All the results come from deterministic selection with seed 42 on test sets of 1,200 samples. The reported result differences are quite small and close to be noise.

Response: The average ACS gains persist across all four states and across repeated TabPFN fits, with small aggregate variation across seeds. Each state has positive mean accuracy/macro-recall increments under both strict and weak disclosure. We additionally reran all 66 tasks with TabPFN seeds 17, 42, 82, and 2026, keeping E0 scores and pseudo-context fixed. All 66 seed-82 runs exactly reproduce the archived predictions. The four-seed mean increments ± sample standard deviations are:

Dataset / conditionDelta accuracy (pp)Delta macro recall (pp)
ACS strict+3.21 ± 0.22+2.09 ± 0.18
ACS weak+2.41 ± 0.01+1.70 ± 0.03
NHIS strict+0.10 ± 0.09+0.62 ± 0.08
NHIS weak+0.85 ± 0.04+1.30 ± 0.03

The ACS accuracy and recall increments remain positive for every tested seed; NHIS macro recall also improves for every seed under both conditions. The ACS mean increments are substantially larger than their variation across backend seeds, supporting reproducible average improvement. Appendix F.3 additionally reports paired bootstrap intervals for all 66 fixed-prediction evaluations. The LLM generation seed is 42 and the tabular seed is 82; Appendix B.3 specifies their roles.

Revised manuscript: Appendix F.3 (existing paired intervals); Appendix B.3 (seed roles). Appendix, Context source, available sample size, and backend seeds (four-seed refits).

Reviewer_Fhw4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes an approach for recovering hidden sensitive attributes in tabular datasets using the remaining available columns (referred to as visible features). The authors investigate two inference approaches: E0 (Row-Level) - An LLM predicts the hidden attribute independently for each record using its visible features; and, E1 (Table-Level) - An enhanced approach that uses the LLM predictions to construct pseudo-labeled examples and train a tabular learner to exploit statistical relationships across the dataset.

The approaches are evaluated using two public datasets, multiple LLMs, and a broad set of LLM-based and non-LLM baselines. Main results are reported mostly for E0.

Critical strengths
  1. The paper demonstrates how pretrained LLM with general knowledge can be used to recover hidden sensitive attributes from visible features, and how these initial predictions can support learning statistical relationships across an entire dataset.

  2. The experimental evaluation includes a broad range of approaches, including traditional ML methods, clustering techniques, and other zero-shot inference methods.

  3. The paper provides clear explanations of the proposed approaches, disclosure settings, and evaluation methodology.

Critical weaknesses
  1. Limited methodological novelty especially taking into account a small performance improvements.

  2. Most of target attributes are represented by relatively simple categorical values, such as four age groups or binary income classes. More challenging attributes, such as occupation and place of birth, achieve considerably lower recovery accuracy, even under the more permissive "weak" disclosure setting. This raises questions about the effectiveness of the approach for recovering more complex or fine-grained private information.

  3. The proposed pseudo-context method (E1) is the main methodological contribution, but its impact is relatively small. For example, the average accuracy improvements over E0 are only 0.22–0.85 percentage points on NHIS. Moreover, the method does not consistently improve predictions across different LLMs, including GPT-4.1. Although detailed results are available in the appendices, the main paper could provide a more critical analysis of when and why E1 improves or degrades inference performance.

  4. In several experiments, the recovery accuracy remains relatively low or is even below the majority-class baseline. Therefore, the practical privacy implications of the reported improvements are not sufficiently convincing.

  5. The experiments consider two survey datasets with overlapped target attributes. It remains unclear whether the approach generalizes to other types of tabular data, more complex sensitive attributes, and datasets with different sizes or statistical characteristics.

Questions for authors
  1. The methodological novelty of E0 and especially E1 appears limited. Personally, I found the comparative evaluations in the appendices, particularly the results across different disclosure settings, datasets, and LLMs, more interesting than the proposed recovery mechanism itself.

I would suggest reconsidering the paper's positioning: rather than introducing another attribute-inference mechanism, the work could emphasize a systematic comparative study of private-attribute inference under different information-disclosure conditions, using both LLM-based and non-LLM approaches. In my opinion, such a study could have greater research impact.

  1. What is the practical significance of the proposed approach, given the relatively small accuracy improvements and the evaluation on mostly categorical attributes? How well would the method perform when recovering more complex sensitive information?

  2. What results could E0 achieve when predicting exact attribute values instead of predefined categories? For example, rather than predicting one of four age groups, could the LLM estimate the person's actual age? What would be the corresponding prediction error?

  3. Although the method is evaluated on two datasets, several target attributes overlap, and both datasets contain survey records. How well would it generalize to other types of tabular data? How does the number of records in the released dataset affect E1's performance?

4 comments
AuthorsOpenReview ↗

Positioning, target complexity, and practical significance (W1-W2, W4, Q1)

We would like to express our sincere gratitude for your careful review and valuable comments and suggestions. Our point-by-point responses are given below.

Weakness 1 / Question 1.

I would suggest reconsidering the paper's positioning: rather than introducing another attribute-inference mechanism, the work could emphasize a systematic comparative study of private-attribute inference under different information-disclosure conditions, using both LLM-based and non-LLM approaches.

Response: We have revised the positioning to foreground a systematic comparative study of zero-shot private-attribute inference, rather than presenting E1 as the sole contribution. The study examines how feature disclosure and aggregate target proportions affect recovery, comparing LLM inference with non-LLM baselines. Its second question is whether the same initial LLM predictions can enable additional recovery through learning across the dataset. E1 provides a way to test this row-level-to-table-level transition. The revised abstract explicitly introduces the systematic study and these two questions; the introduction and Section 5.2 organize the contributions and evidence accordingly. The pseudo-context strategy remains a method contribution within this broader investigation.

Revised manuscript: Abstract; Section 1 (contributions); Section 5.2.

Weakness 2.

More challenging attributes, such as occupation and place of birth, achieve considerably lower recovery accuracy, even under the more permissive "weak" disclosure setting. This raises questions about the effectiveness of the approach for recovering more complex or fine-grained private information.

Response: OCC and POB are more difficult to recover than the simpler targets, but the results still demonstrate useful inference on these larger label spaces. On ACS weak disclosure, averaged across four states, E1 improves both accuracy and Macro-F1 for the 12-class OCC and 20-class POB targets:

TargetAccuracy: E0 → E1Macro-F1: E0 → E1
OCC39.98% → 40.48%11.80% → 12.41%
POB24.71% → 25.17%4.56% → 4.69%

OCC macro recall also rises from 17.16% to 17.58%, versus 8.33% for Train Majority. The recovery therefore extends beyond predicting the dominant class. These gains use unchanged initial LLM outputs, demonstrating additional attribute recovery even for the larger target spaces.

The candidate rankings also retain information that top-1 accuracy does not capture. For ACS weak OCC, the correct label appears among the first five candidates for 73.08% of samples, compared with 39.98% top-1 accuracy (Appendix E.4, Figure 4). This measures candidate coverage rather than final E1 accuracy, but shows that the LLM often identifies a useful set of alternatives even when its first choice is incorrect. Section 6 proposes broad-to-fine inference for occupation and birthplace as a future extension for finer target resolutions.

Revised manuscript: Table 3 (accuracy and macro recall); Appendix E.4, Figure 4 (candidate coverage and additional recovery metrics); Section 6 (hierarchical extension).

Weakness 4.

In several experiments, the recovery accuracy remains relatively low or is even below the majority-class baseline. Therefore, the practical privacy implications of the reported improvements are not sufficiently convincing.

Response: Train Majority is a reference with additional label access. It identifies the most frequent value from true training labels, which are unavailable to the strict/weak attacker. Treating it as a minimum success threshold therefore compares unequal information access. A common-value guess from general knowledge remains possible, but does not supply the empirical training majority.

Recovery must also be assessed across classes. On ACS education under weak disclosure, E0 accuracy is below Train Majority (34.02% versus 44.52%), but macro recall is 41.97% versus 20.00%, and within-one-level accuracy is 90.77% versus 80.46%. It recovers more than the dominant class and often stays close to the true level. These are attributes the publisher intended to withhold.

The appropriate incremental comparison is E1 versus E0: both have the same information access and initial LLM outputs. This measures how much additional recovery comes from using information across samples.

Revised manuscript: Section 5.2 (Majority and metric comparisons); Section 6 (interpretation of recovery).

AuthorsOpenReview ↗

Additional recovery, practical significance, and model variation (W3, Q2)

Weakness 3.

The method does not consistently improve predictions across different LLMs, including GPT-4.1. Although detailed results are available in the appendices, the main paper could provide a more critical analysis of when and why E1 improves or degrades inference performance.

Response: We agree that E1's effect varies across datasets and LLMs, and have brought the complete matched comparison into Section 5.2. The analysis now asks what distinguishes net improvement from net degradation, using all 24 strict/weak tasks shared by Qwen and GPT: seven ACS NY targets and five NHIS targets under both conditions. Table 4 reports the following mean increments over each model's own E0 (percentage points):

Dataset / conditionQwen accuracy / macro recallGPT accuracy / macro recall
ACS NY strict+3.95 / +2.90+0.92 / +1.86
ACS NY weak+2.42 / +1.64-0.63 / +0.57
NHIS strict+0.22 / +0.71-0.13 / -0.28
NHIS weak+0.85 / +1.32-1.15 / -0.67

Further recovery remains possible after a stronger E0. Of the 2,064 errors corrected by Qwen E1, 1,228 (59.50%) were already correct under GPT E0, leaving GPT less opportunity for further recovery. Nevertheless, E1 raises mean GPT accuracy across the seven ACS NY strict targets from 47.32% to 48.24%, with macro recall increasing by 1.86 points. ACS NY weak macro recall also increases by 0.57 points. These are additional recoveries over GPT's own predictions with the initial scores held fixed.

What determines improvement or degradation? The tabular learner treats selected pseudo-labels as supervision. When the selected examples capture useful feature--target relationships, transfer can correct E0 errors; when they carry systematic labeling mistakes, the learner can propagate those mistakes to initially correct samples. Stronger E0 alone explains reduced headroom, not a net decrease. Selection must provide useful pseudo-supervision relative to that stronger starting point.

The matched diagnostics support this distinction. For NHIS weak SCH, GPT context precision is 38.80%, versus 38.92% E0 accuracy, with four of five classes represented. Selection does not improve label correctness in this case. This is not a claim that GPT pseudo-labels are uniformly worse: GPT context precision exceeds Qwen's on 14/24 tasks. Tables 26-27 therefore assess context quality alongside correction and error propagation. They also locate the loss before final fusion: GPT fused accuracy exceeds direct TabPFN accuracy in all four groups. Thus the observed decreases cannot simply be assigned to the product-fusion rule.

The revised analysis distinguishes initial recovery, usefulness of the selected supervision, and subsequent transfer. It preserves the complete matched comparison while showing that additional private-attribute recovery remains possible with the stronger commercial model.

Revised manuscript: Section 5.2 and Table 4 (complete matched comparison); Tables 26-27, Appendix F.2 (context quality, corrected/introduced errors, and direct/fused endpoints).

Question 2.

What is the practical significance of the proposed approach, given the relatively small accuracy improvements and the evaluation on mostly categorical attributes? How well would the method perform when recovering more complex sensitive information?

Response: The practical implication is that the released dataset can strengthen an attack beyond its initial LLM predictions. On ACS weak, E1 corrects 2,184 E0 errors and introduces 1,366, yielding 818 net additional correct attribute predictions across 33,600 sample-target evaluations (Table 25). It uses the existing LLM outputs without further LLM calls or ground-truth target labels. Income brackets and education levels can themselves be information the publisher intended to withhold, so categorical recovery also has privacy relevance.

The method also accommodates structured targets: our LOC extension uses coordinate regression to inform candidate reranking, improving weak-disclosure Acc@100km from 39.06% to 41.08%. For finer occupational or birthplace values, Section 6 proposes scoring broad groups before finer candidates and using the resulting predictions as pseudo-supervision. The new exact-age experiment in our Q3 response additionally evaluates numerical recovery in years using regression-based pseudo-supervision; the hierarchical extension remains a direction for finer occupational and birthplace recovery.

Revised manuscript: Section 5.2 (closed-set and LOC recovery); Table 25, Appendix F.2 (net correction counts); Section 6 (hierarchical and numerical extensions).

AuthorsOpenReview ↗

Exact age, domain scope, and dataset size (W5, Q3-Q4)

Question 3.

What results could E0 achieve when predicting exact attribute values instead of predefined categories? For example, rather than predicting one of four age groups, could the LLM estimate the person's actual age? What would be the corresponding prediction error?

Response: We supplemented the grouped-age evaluation with an exact-age experiment on 1,200 NY samples under weak disclosure. This supplementary experiment predicts integer ages rather than one of four age groups; its configuration is specified in the appendix. The following comparison uses the same five-candidate LLM outputs for E0 and E1:

MetricE0E1
Mean absolute error (years) ↓22.1621.94
Median absolute error (years) ↓1716
Root mean squared error (years) ↓29.0528.79
Exact-age accuracy ↑2.67%2.92%

For numerical E1, background-adjusted evidence within ten-year strata selects 305 exact-age pseudo-labels for a TabPFN regressor, whose numerical prediction informs candidate reranking. The strata are used only for context selection; predictions and evaluation retain individual years. E1 reduces mean and median age errors and increases exact-age accuracy, extending the pseudo-context procedure to a numerical target.

Revised manuscript: Appendix, "Supplementary exact-age inference" (new numerical-age experiment and configuration).

Weakness 5 / Question 4.

Although the method is evaluated on two datasets, several target attributes overlap, and both datasets contain survey records. How well would it generalize to other types of tabular data? How does the number of records in the released dataset affect E1's performance?

Response: We added a Qwen3-1.7B financial-domain experiment on all 1,000 Statlog German Credit applications, evaluating checking-account status (four classes) and savings-account status (five classes). Both tasks hide the two account-status columns and the credit-risk outcome, retaining 18 visible features such as loan amount/duration, credit history, employment, property, housing, age, and job. This financial-attribute-hidden policy is separate from ACS strict/weak. The attacker receives zero ground-truth target labels and no class proportions. The same closed-set E1 selects 68 checking and 186 savings pseudo-labeled context samples, followed by TabPFN and product fusion.

TargetAccuracy: E0 / E1Macro recall: E0 / E1
Checking7.40% / 12.20%25.36% / 23.71%
Savings5.80% / 6.60%18.73% / 20.60%

E1 improves accuracy on both financial targets, by 4.80 and 0.80 percentage points respectively, and also improves savings macro recall. This extends the test of pseudo-context transfer beyond survey attributes. Configurations and results, including the separate Qwen3-14B financial run, are provided in the appendix.

We have now completed a controlled scale study using cached LLM outputs. Each of the 66 strict/weak tasks uses 200 fixed queried samples and nested donor pools of 100, 200, 400, 800, and 1,000 other samples. Background scores and context are recomputed at each size; query samples never enter context. Three predetermined random partitions give 990 runs. Mean ACS accuracy increments over the same query E0 predictions are:

Donor samplesStrict ΔAccuracy (pp)Weak ΔAccuracy (pp)
100+1.15+1.39
200+1.61+1.86
400+2.48+1.90
800+2.92+2.16
1,000+3.18+2.18

ACS macro-recall increments also increase from approximately +0.00 to +1.85 points under strict and from +0.08 to +1.48 under weak. NHIS weak accuracy changes from −2.00 to +0.53 points, while NHIS strict is non-monotonic. Thus, more available samples strengthen average recovery in these ACS comparisons, but not necessarily in every population or task. This measures donor availability separately from selected context size. It is a fixed-query diagnostic within the existing cache, rather than a new-domain evaluation or the original full-table endpoint.

Revised manuscript: Appendix, "Supplementary financial-attribute inference" (new domain experiment and configuration). Appendix, Context source, available sample size, and backend seeds (size controls).

I appreciate the authors' efforts to improve the paper's clarity and provide additional experiments. However, the recovery performance remains limited and the additional results do not sufficiently strengthen the paper's practical significance (for instance, the recovery of exact values). Therefore, I am keeping my original assessment.

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

This work studies zero-shot attribute inference on anonymized tabular cohorts by distilling world knowledge from pretrained LLMs into a downstream tabular learner. Specifically, the authors use row-wise zero-shot LLM confidence scores (E0E_0) as heuristics to select and pseudo-label a context subset, which then trains a transductive tabular model to refine predictions across the table (E1E_1). While the empirical benchmark spans multiple targets, LLM backbones, and feature/prior disclosure levels, the core finding that suppressing sensitive columns fails to prevent inference attack when correlated quasi-identifiers or marginal priors are exposed, is well established in the privacy literature, and the proposed pseudo-labeling pipeline yields marginal or negative gains when paired with stronger LLM backbones.

Critical strengths
  • By freezing the row-level LLM score matrix across both E0E_0 and E1E_1, the experimental design cleanly isolates the marginal gain of transductive tabular learning without confounding it with extra LLM inference passes. This provides a clean diagnostic framework for evaluating semi-supervised transfer from foundation models.
  • Evaluating four LLM families alongside prior-assisted tabular clustering and zero-shot prototype/tree baselines across 13 ACS/NHIS targets provides a broad empirical reference point for how pretrained language priors compare against unsupervised tabular structure when ground-truth labels are absent.
Critical weaknesses
  • Suppressing sensitive columns while exposing correlated quasi-identifiers has been known to be vulnerable for privacy attacks, and zero-shot LLM inference on ACS rows was already shown by Staab et al. (2024). Evaluating attacks solely via classification accuracy and macro recall conflates population-level statistical imputation with dataset-specific privacy leakage (Jayaraman & Evans, 2022), especially without measuring adversarial advantage over a distributional baseline or formal privacy bounds.
  • The proposed E1E_1 simply combines zero-shot LLM pseudo-labeling with TabPFN self-training, and its empirical benefits do not hold up across datasets or stronger models, such as GPT-4.1.
  • The authors claimed that model checkpoints predate the 2024 ACS and 2024 NHIS releases only reduced exact record memorization. Because ACS PUMS and NHIS are annual surveys with virtually unchanged variable codebooks and joint demographic distributions across years, the evaluated LLMs may have already memorized these exact survey correlations from pre-2024 releases during pretraining.
Questions for authors
  1. How do the authors distinguish population-level statistical imputation from true privacy leakage caused by cohort release (Jayaraman & Evans, 2022), and why does E1E_1 degrade GPT-4.1 accuracy across most disclosure settings (Tables 17–18) if the cohort genuinely provides additional private signal?
  2. Can the authors evaluate E0E_0/E1E_1 on domain-shifted or private tabular datasets that do not have pre-2024 annual counterparts in public LLM pretraining corpora?
  3. Since E1E_1 only reranks the k=5k=5 cities generated by E0E_0 truncating support for non-hub locations and weights context selection by spatial concentration, why should LOC be considered a successful attack when both E0E_0 and E1E_1 lose to Train Majority based on results from Tables 3 & 10?
  4. What principled threat-model criterion governed the feature partition in Table 5, given that strict still exposes strong correlates, such as industry INDP and work hours WKHP when inferring occupation or income?
3 comments
AuthorsOpenReview ↗

Privacy distinction and feature partition (W1, Q1, Q4)

We would like to express our sincere gratitude for your careful review and valuable comments and suggestions. Our point-by-point responses are given below.

Weakness 1 / Question 1 (privacy distinction).

How do the authors distinguish population-level statistical imputation from true privacy leakage caused by cohort release (Jayaraman & Evans, 2022)?

Response: Staab et al.'s ACS experiment corresponds to the row-level inference setting represented by our E0: an LLM infers a hidden attribute from the visible information of each sample independently. We build on that result. Our question is whether the same initial predictions can support additional table-level inference, by providing pseudo-labels for learning dependencies across samples. With the LLM outputs held fixed, E1 increases mean accuracy by 3.35 and 2.43 percentage points across seven ACS targets under strict and weak disclosure. The study also measures how revealing other private attributes or aggregate proportions affects recovery.

The central comparison in Jayaraman and Evans (2022) is inference with versus without access to a model trained on sensitive data, while controlling the adversary's auxiliary knowledge. Their imputation experiments learn from auxiliary samples whose sensitive attributes are available, with varying sample sizes and distributional similarity. Our adversary instead receives no ground-truth target labels for fitting or tuning the attack. We do not assume access to a publisher's model trained on the sensitive dataset. The attacker uses general pretrained models, and the supervision for the subsequent tabular learner comes entirely from LLM predictions.

Thus, the difference concerns both the available resource and the comparison: their study isolates information obtained from a sensitive-data-trained model; ours measures the additional recovery obtained by using zero-shot predictions for statistical learning across an unlabeled table. Pretrained distributional knowledge is permitted and is already used by E0. Our supervised references assess the stronger-access case where genuine target labels are available.

Revised manuscript: Sections 3.1, 3.2 and 4.1; Section 6 (privacy interpretation).

Question 4.

What principled threat-model criterion governed the feature partition in Table 5, given that strict still exposes strong correlates, such as industry INDP and work hours WKHP when inferring occupation or income?

Response: Following Staab et al.'s choice of personal-attribute categories, we distinguish the private attributes to be withheld from the background features available to the attacker. The background features retained under strict are assumed to be non-sensitive and publicly available within our threat model; this is why they are retained, even when they are strongly related to a private target. This assumption specifies their role in our disclosure policy, rather than asserting that these features are non-sensitive in every application. Strict withholds the fields defining the designated private targets; weak reveals selected other private attributes while keeping the attacked target hidden. This allows us to test whether exposing some private attributes makes the remaining ones easier to infer.

Feature retention is governed by the assumed privacy policy, not by the absence of predictive dependencies. We study whether information treated as non-sensitive under that policy can support inference of a concealed private attribute. Removing all dependencies would remove the information needed for sample-specific inference and would answer a different question from whether withholding the designated private columns is sufficient protection.

If the visible features contained no information about the target, sample-specific prediction could not improve on a predictor using only the target's marginal distribution. The relevant privacy question is consequently whether the retained background information supports recovery of an attribute that the publisher intends to hide. Our feature policy makes this question explicit, while the feature table documents its implementation.

Revised manuscript: Section 3.2; Table 6, Appendix A.3.

AuthorsOpenReview ↗

E1 and GPT-4.1; LOC (W2, Q1, Q3)

Weakness 2 / Question 1 (GPT-4.1).

Why does E1 degrade GPT-4.1 accuracy across most disclosure settings (Tables 17-18) if the cohort genuinely provides additional private signal?

Response: GPT-4.1 provides stronger initial predictions and therefore leaves less recovery to the table-level step. In our 24 matched strict/weak tasks, 59.50% of the errors corrected by Qwen E1 were already correct under GPT E0. Nevertheless, E1 still recovers additional private attributes beyond GPT's own row-level predictions. This additional recovery is the relevant test of whether the table can provide useful information after a strong LLM has made its initial predictions.

Additional recovery with GPT-4.1. Across the seven ACS New York targets under strict disclosure, E1 raises GPT's mean accuracy from 47.32% to 48.24% (+0.92 percentage points). At the target level, INC improves from 73.17% to 79.08% (+5.92 points), and SCH from 38.58% to 40.83% (+2.25 points). On NHIS strict AGE, accuracy improves from 40.08% to 43.83% (+3.75 points). These are net improvements over GPT's own E0, after accounting for both corrected predictions and newly introduced errors. They use the same initial GPT outputs, no ground-truth target labels, and no additional LLM calls: the extra recovery comes from constructing pseudo-context and applying tabular inference.

Why some group averages decrease. E1 fits pseudo-labels as supervision. Their mistakes can therefore be learned as feature dependencies and propagated to samples whose E0 predictions were correct. Stronger E0 reduces correction opportunities but does not ensure that selection produces better training labels. For NHIS weak SCH, GPT context precision is 38.80%, versus 38.92% E0 accuracy, with only four of five classes represented: selection supplies noisy, incomplete supervision. The loss arises in tabular transfer, not simply fusion: GPT fused accuracy exceeds direct TabPFN accuracy in all four matched groups, while both remain below E0 in the negative groups (Appendix F.2, Tables 26-27). These comparisons are consistent with pseudo-label error propagation.

These results show that table-level inference can recover additional private attributes even after GPT-4.1's stronger initial predictions. Section 5.2 reports the matched comparisons, and Appendix F.2 analyzes the context quality and prediction changes associated with the different outcomes.

Revised manuscript: Section 5.2 (matched-model results); Appendix F.2 (additional recovery, context quality, and prediction changes), Tables 26-27.

Question 3.

Since E1 only reranks the k = 5 cities generated by E0, truncating support for non-hub locations, and weights context selection by spatial concentration, why should LOC be considered a successful attack when both E0 and E1 lose to Train Majority based on results from Tables 3 & 10?

Response: Train Majority is a label-dependent reference, rather than an attack under the same zero-shot information constraints. It uses the true training labels to identify the most frequent location. Neither those labels nor their empirical location proportions are supplied to E0/E1 under strict/weak disclosure. Consequently, directly treating this reference as the minimum performance required of a zero-shot LOC attack is not a comparison under equal information access. An attacker can guess a common location from general knowledge, but identifying the empirical majority from labeled training data is an additional resource.

The LOC result establishes additional recovery relative to the same attacker's E0. On the four-state Qwen weak experiment, E1 increases accuracy within 100 km from 39.06% to 41.08% and reduces median error from 162.74 to 156.24 km, using the same initial candidates and scores.

Recovery beyond the constant prediction across all four states. For all 4,800 Qwen weak LOC samples, the following counts use the same 100 km success criterion:

StateE1 hits missed by MajorityAlso missed by E0
CA12237
IL3733
NY2912
TX7941
Total267123

Thus, E1 recovers locations missed by Majority in every state, including 123 also missed by E0. Over the full sample, E1 corrects 430 E0 errors and introduces 333, giving 97 net additional hits. Appendix E.4 reports the candidate coverage that bounds reranking.

Revised manuscript: Section 5.2 and Appendix E.1 (LOC results); Appendix F.2 (full-test-set LOC recovery counts).

AuthorsOpenReview ↗

Pretrained knowledge and additional domains (W3, Q2)

Weakness 3 / Question 2.

The evaluated LLMs may have already memorized these exact survey correlations from pre-2024 releases during pretraining.

Can the authors evaluate E0/E1 on domain-shifted or private tabular datasets that do not have pre-2024 annual counterparts in public LLM pretraining corpora?

Response: We agree that model-before-release timing does not exclude knowledge of earlier survey schemas or population correlations. We have narrowed that statement accordingly. Pretrained population knowledge is an allowed resource in our threat model and supplies E0's initial predictions. We choose evaluated annual releases published after the selected model checkpoints to reduce the possibility of pretraining access to those particular data. The intended resource is the general attribute relationships and population patterns learned during pretraining, which the attacker applies to newly released samples. This timing reduces potential direct exposure without proving that the model has never encountered related data.

To examine transfer across different population sources, we have added a cross-state context experiment, also discussed in our response to Reviewer QXPp. Each of the four ACS states is evaluated using pseudo-context from each of the other three states, for all seven closed-set targets under strict/weak disclosure: 168 source-destination/task combinations. The destination samples and E0 scores remain fixed; the source context uses LLM pseudo-labels and no ground-truth target labels. Mean increments over destination E0 are:

Context sourceStrict accuracy / macro recall (pp)Weak accuracy / macro recall (pp)
Destination's own context+3.35 / +2.20+2.43 / +1.73
Another state's context+3.04 / +1.88+2.18 / +1.39

The external-context gains also persist when evaluation is restricted to the identical destination samples outside its original selected context: +3.50/+2.36 points under strict and +2.96/+1.56 under weak. Thus pseudo-supervision can transfer between state populations; the additional inference does not require context from the attacked individuals themselves.

This control demonstrates cross-state transfer within ACS under a shared schema. Fully excluding pretraining exposure to similar data patterns is difficult. Exploring whether targeted machine unlearning could help disentangle this prior knowledge would require a separate study beyond the scope of this paper.

Revised manuscript: Sections 5.1.1 and 6; Appendix, Context source, available sample size, and backend seeds (cross-state controls).

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

Related papers

Discussion (0)

Sign in to comment.