acceptodds
Under review as a conference paper at ICLR 2027

HashMark: Type-Agnostic Watermarking for Tables with Free-Form Text

Abstract

Tables are released, copied, and edited far from their owners. A copy found later may have had its rows, values or text changed. We introduce HashMark, a watermark for tables containing both numbers and free-form text. It makes small numerical edits and selects text paraphrases under one secret key. Verification requires only the suspect table, the key, and a public descriptor of how it was marked; neither the original table nor its generator is needed. Column-level statistical tests account for repeated values and dependence between text cells, providing evidence that a suspect table contains watermarked data. On mixed tables, the numbers and the text each keep the verdict when the other is attacked, and no clean control or wrong key is flagged in any mixed-table test. With of the rows or cells of a table replaced, HashMark still detects every table, while the generation-time watermark MUSE detects –. A geometric model predicts how many text cells can be marked to within percentage points across four corpora. Numerical marking keeps the data as close to the original as an unwatermarked release, to three decimal places, moves downstream accuracy by at most points, and marks rows in  s.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
watermarking, tabular data, synthetic data, cryptographic hash function, robust

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

This paper proposes HashMark, a watermarking framework for tabular data containing numerical values and free-form text. It embeds watermarks through keyed numerical modifications and semantic paraphrase selection, enabling blind detection without the original data or generative model. Experiments demonstrate robustness against various data modification attacks while preserving data utility.

Critical strengths

1.Practical applicability: HashMark supports both numerical and free-form text data, enabling blind watermark detection without access to the original data or generative model.

2.Theoretical analysis: The paper provides comprehensive mathematical analyses of watermark capacity, distortion bounds, and statistical detection.

3.Presentation: The presentation is clear.

Critical weaknesses
  1. Limited methodological novelty: The core watermarking mechanisms of HashMark are closely related to existing approaches. Its numerical component relies on keyed hashing and small value perturbations share similar principles with prior tabular watermarking methods[1,2,3]. Its text watermarking component largely follows the semantic-space partitioning and rejection-sampling paradigm introduced by SemStamp[4]. A comprehensive comparsion and discussion should be added to clarify its novelty and improvements beyond integrating existing techniques.

  2. Robustness: The numerical watermark relies on hashing values at a predefined decimal precision, making it inherently sensitive to small numerical perturbations. In practical data processing pipelines, rounding or numerical formatting, precision conversion is common. This limits the practical applicability of the proposed method. Appendix L.1 shows the strong fidelity and robustness results of HashMark are achieved under different operating conditions rather than simultaneously. A more systematic comparison of the fidelity–robustness trade-off would strengthen the paper.

  3. Semantic Fidelity Concerns:The proposed text watermarking method selects paraphrases according to their positions relative to a secret hyperplane rather than explicitly optimizing semantic equivalence with the original text. These findings raise concerns about the the method for applications requiring precise preservation of factual information, such as medical records. The paper already include evaluates semantic fidelity using automated language models, but additional human evaluation or stronger semantic consistency constraints would be valuable.

[1]He, Hengzhi, et al. "Watermarking generative tabular data." arXiv preprint arXiv:2405.14018 (2024).

[2]Zheng, Yihao, et al. "Tabularmark: Watermarking tabular datasets for machine learning." Proceedings of the 2024 on ACM SIGSAC conference on computer and communications security. 2024.

[3]Ngo, Dung Daniel, et al. "Adaptive and robust watermark for generative tabular data." arXiv preprint arXiv:2409.14700 (2024).

[4]Hou, Abe, et al. "Semstamp: A semantic watermark with paraphrastic robustness for text generation." Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Questions for authors

See weakness. I am willing to revise my score based on the authors' response.

3 comments

We thank the reviewer for their comments. W1 (novelty against [1–4]). HashMark keeps SemStamp's idea that a secret hyperplane in embedding space can carry a watermark. SemStamp's construction, however, relies on a sequence: each sentence's valid region is seeded by the previous sentence, and the language model is resampled until the next sentence lands in it. A table has no such sequence. Its cells are independent records, rows can be sorted or deleted, and a cell is often a single sentence. HashMark therefore paraphrases each cell on its own and keys every cell in a column to one hyperplane, so a cell's validity depends only on its own text and the column descriptor. This is simpler and order-free, but raises two problems SemStamp does not face.

Dependence between cells. With one shared hyperplane, similar cells land on the same side together, and sentence embeddings point in a common direction. An unmarked column can then show a large excess of valid cells: in our uncentered adaptation, clean scores have standard deviation 10.55 where an independent-bit test assumes 1 (Table 8), and a margin rule leaves it at 10.86 (Table 24). HashMark centers the embeddings on the column's published mean, derives the null variance from the pairwise angular correlations (Proposition 1), and decides ownership against fresh keys. Centered bits are still dependent (null standard deviation 0.71), so a raw z score is not a standard-normal p-value.

Capacity under a finite paraphrase budget. Meaning-preserving paraphrases of a cell cluster in embedding space, so extra candidates are not independent chances to cross the hyperplane, and some cells stay unmarkable. Proposition 2 predicts this failure probability from the candidates' geometry, and hence the expected valid fraction and score, with no fitted parameters. Appendix E.2 shows that a pilot pool of four paraphrases per cell already predicts the capacity of a pool of 32.

Numerical columns. WGTD [1], TabularMark [2] and Ngo et al. [3] move values into keyed intervals or sub-ranges. HashMark's Edit Mode instead searches the declared precision grid for the first valid point and leaves repeated inputs unchanged. Theorem 1 bounds both the largest change and the shift in each column's distribution. Detection counts each distinct value once, since repeats carry no new hash bit; without this, clean scores have standard deviation 22.6 instead of 1 (Appendix D.1). Coarser rounding than declared and noise need a different operating point (W2).

One decision for the table. HashMark declares the table marked if any column passes its corrected threshold, needing only the key and the published descriptor; prior tabular schemes need the original data, a generator at marking time, or a sampling protocol over candidate rows. The family-wise decision needs no independence between columns: on Airbnb, rounding away the numerical evidence leaves the text to decide, rewriting away the text leaves the numbers to decide, and only both attacks together remove the verdict (Table 17).

Why the combination needs these changes. The composed baseline combines WGTD-style binning and SemStamp-style selection under our table protocol, without centering or distinct-value counting. On an unwatermarked Airbnb table it scores 35.5, against 1.3 for HashMark. It separates marked from clean tables by 2.8 null standard deviations, against 67.8 for HashMark, and loses its evidence under rounding (Appendix F.1.2, Table 18). The revised related work states the inherited ingredients and these contributions.

Settings supported by the closest methods (revised Table 1):

WGTD [1]TabularMark [2]Ngo [3]MUSESemStamp [4]HashMark
Datanumbersany, text untestednumbersgenerated rows, text untestedgenerated textnumbers and free-form text; categories unchanged
Blind detectionyesnoyesyesyesyes
Marks a finished tableyesyesyesnonoyes
One decision across numbers and textnononononoyes

Comparison with MUSE and Ngo et al. HashMark marks 10,000 numerical rows in about 0.2 seconds, where MUSE samples several candidate rows for each released row. On Shoppers, after 80% row or cell replacement, HashMark detects every tested table, against 26% and 34% for MUSE (Table 33). MUSE is more robust under value alteration and on Beijing (Tables 33 and 37). On Adult, HashMark's normalized C2ST is 0.996, equal to the unwatermarked release, against 0.883 for MUSE (Table 3). Our reimplementation of Ngo et al. (Appendix J.2), on Magic where its test is calibrated, survives shuffling and moderate row replacement, but detects at most 10% of tables at 80% value alteration and 1% after rounding to two decimals. On the same tables, Edit Mode detects 100% under value alteration, and HashMark-Q with 32 bins detects 100% after rounding, with a threshold calibrated in-sample.

W2 (rounding, and fidelity with robustness). The strongest fidelity result for Edit Mode and the strongest noise-robustness result for HashMark-Q describe different operating points. The revision adds a more systematic comparison by bin count (Table 44, Appendix L.1).

For Edit Mode, the owner declares a precision c and the attacker knows it. Rounding to c or more decimal places preserves the released grid values and hence the watermark. Rounding one decimal place more coarsely removes the exact-value evidence in Figure 2. For latitude at c = 6, that coarser scale is approximately 1.1 metres; the declared grid spacing itself is approximately 11 cm. Robustness to formatting at the declared precision does not extend to arbitrary numerical perturbations.

HashMark-Q trades larger changes for robustness by marking bins rather than exact values. There are operating points with both useful fidelity and rounding robustness: with 512–1024 bins, it retains z = 20–22 after rounding to two decimals, while a classifier distinguishes marked from unmarked rows with accuracy 0.51–0.55, close to chance.

Substantial multiplicative noise is more costly. At noise strength α = 0.2, 32 or fewer bins are needed to detect at least 99% of Shoppers tables; exact Edit Mode detects none (Table 31), and 32 bins detect 90% on Magic (Appendix J.2). The reported classifier accuracy at the 32-bin operating point is 0.898, so the perturbation is distinguishable. Table 44 shows classifier accuracy falling from 0.94 at 8 bins to 0.64 at 256 bins. The owner therefore chooses the operating point for the expected attack.

W3 (semantic fidelity). The hyperplane criterion does not ensure factual equivalence. A candidate can carry the mark and still be a poor rewrite. We have added human evaluation and experiments with factual-preservation instructions to examine this concern more directly (Appendix K.1).

First, we use DeBERTa-large-MNLI, independently of the watermark encoder, to test entailment in both directions between the original and released cells. Across 500 cells, HashMark leaves 335 unchanged and achieves a bidirectional-entailment rate of 0.746, compared with 0.282 when every cell is paraphrased. That aggregate result benefits substantially from leaving cells unchanged. Among the 165 rewritten cells alone, the keyed choice has a lower entailment rate than a random paraphrase, 0.230 versus 0.297, and a higher contradiction rate, 18.2% versus 7.9%. These results support your concern that selection can worsen an already imperfect candidate pool.

Second, a blinded annotator who was not involved in developing the method rated meaning preservation from 1 to 5 on 156 cells rewritten by HashMark. The mean was 2.92 for the keyed choice and 2.99 for a random paraphrase, a difference of −0.07 with 95% confidence interval [−0.19, 0.05]. The annotator found contradictions in 3.8% of cells in both conditions. This evaluation does not show a clear additional loss from keyed selection, but the absolute scores leave room for improvement, and one annotator provides limited evidence about precise factual preservation.

Finally, we tested 500 Drugs.com patient reviews. An LLM judge found that only 3% of the T5 rewrites HashMark released (and 1% of T5 paraphrases of the cells the local model rewrote) preserved every fact, and DeBERTa filtering did not adequately identify the failures. We therefore generated candidates with an LLM explicitly instructed to retain every fact, stopping at the first candidate on the valid side of the hyperplane. The rewritten reviews retained 96% of drug names and doses, compared with 72% for T5 rewrites of the same reviews; this particular check counts the names and doses without a model. Detection remained strong at z = 10.9. With a locally run 7B model, an independent judge rated 51% of rewrites as preserving every fact, compared with 1% for T5, while the watermark remained detectable at z = 8.2.

Better candidates therefore improve factual preservation and keep the watermark. Text whose exact facts matter should be rewritten only by a generator instructed to keep them, and its rewrites validated independently.

Reviewer_cWkrOpenReview ↗

Thank you for your reply

Thank you for the detailed response and the additional experiments. The revision addresses some of my concerns.

  1. The revision clarifies the contribution of adapting existing watermarking mechanisms to unordered mixed tables. I still see the novelty as modest, and the composed baseline leaves open how much HashMark improves over a stronger combination with separately calibrated numerical and text detectors.

  2. Table 44 addresses my request for a systematic comparison. The revision now makes clear what fidelity cost accompanies each robustness result.

  3. Semantic fidelity remains my main concern. Subtle numerical changes are hard for human raters to judge, so quantitative distortion and downstream utility are more informative for numerical columns. For rewritten text, the blinded evaluation is relevant, but the mean meaning-preservation score of 2.92 out of 5 leaves my concern about factual fidelity unresolved.

I really appreciate the time and effort the authors have put into the rebuttal and additional experiments. After careful consideration, I've decided to keep my original score. Thanks again to the authors for their efforts and thoughtful responses!

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

This paper proposes HashMark, which is a tabular watermarking method that embeds both numerical and free-form text columns using a single secret key. It detects the embedded watermark without requiring original tabular data or generator. The paper provides a light theoretical analysis and experiments against recent tabular watermarking methods and shows strong watermark detectability. The paper also conducts comparisons against several post-editing attacks, some of them remain detectable, and some of them are not.

Critical strengths

Generator agnostic verification that covers mixed data types. It detects a suspect table using only the key and a public descriptor. The embedding is cheap (about 0.2sec per 10k rows of numbers) and survives under sorting and subsampling. This makes watermarking practical in real world applications.

Critical weaknesses
  • Novelty and several headline claims are overstated. Those components are already known, such as keyed hash value marking and SemStamp’s style hyperplane selection with centering, combined by Bonferroni. The contribution is mainly an integration of existing components into a tabular watermarking method.
  • The main text is hard to follow because key evidence are deferred to the appendix. Some numbers are TPR@1%FPR, some use the z > 3.09 reference, and some are descriptive z/SD ratios that Table 3’s caption warns are not a p-value. These metrics are not well presented or organized, and many of the headline evidence are only in the appendix.
Questions for authors
  • Figure 3 shows the predicted lines are computed from the same paraphrase pools that the measured markers use. Does the prediction use any information beyond the realized pools? If not, can it estimate capacity before the full pools are generated, for example from a few paraphrases per cell?
  • Page 8 mentions that replacing each summary with another of the owner’s paraphrases erases the text mark. However, the paraphraser is public, so can an attacker regenerate the owner’s candidates, at least for the cells released unchanged? Is the generation seed part of the secret key?
  • How many independent clean and wrong key controls support the claim that none of them is flagged in any mixed table test, and what upper confidence bound on the false positive rate do they give? The ethics statement mentions that false positive rates vary across fixed keys. Does the decision rule’s guarantee hold for an innocent table scored under the owner’s deployed key?
2 comments

We thank the reviewer for the detailed comments. We hope the reviewer will consider our comment when updating their assessment.

Novelty. Keyed hashing, SemStamp's hyperplane selection, and Bonferroni correction are established tools; centering is our addition. Our contribution is a simple construction for finished tables whose numerical and text cells can be checked without the original data or generator, along with the analysis needed to make detection meaningful. Repeated numerical values and correlated sentence embeddings invalidate a naive independent-bit test: the composed baseline gives an unwatermarked Airbnb table a score of 35.5, compared with 1.3 for HashMark (Appendix F.1.2).

Our reply to cWkr (W1) gives the full comparison: what we retain from SemStamp, why its sequential construction does not carry over to table cells, the numerical construction against prior binning methods, and results against MUSE and Ngo et al.

Metrics. Table 2 now identifies the detection rule and metric for every experiment. Ownership claims in Table 4 and Tables 16–17 use the calibrated family-wise decision. The reported p-value belongs to the deciding column: a numerical tail probability or a fresh-key text p-value. The MUSE attack grids (Tables 33 and 35–37) instead report TPR at 1% FPR, following that benchmark, with thresholds estimated from attacked clean tables. Raw z-scores in Figures 2 and 4 describe how evidence changes with an attack; their reference line at 3.09 only helps compare curves, and ownership is decided by the calibrated test. We moved the z/SD summaries to Table 14 and labeled them as descriptive.

Q1 (capacity prediction). Figure 3 uses the realized candidate pools: Proposition 2 computes, from their embedding geometry alone, the probability that all candidates fall on the invalid side of a random hyperplane, with no fitted parameters. Appendix E.2 now tests a small pilot budget. Four paraphrases per cell estimate the exhaustion rate for a pool of 32 within 1.1–2.0 percentage points across four corpora, overestimating exhaustion in every case; the predicted z is within 0.9. A small pilot pool can therefore guide the generation budget. The geometry-based prediction also transfers to an unseen encoder and paraphraser, with exhaustion error within 0.0045 (Table 13). The prediction needs the pools to be independent of the hyperplane, which holds both for our public seeds and for the owner's secret randomness in deployment (Q2).

Q2 (regenerating the owner's candidates). The paraphraser is public, but the owner's generation randomness is secret. It is sampled independently of S and never published; the detector needs only S and the public descriptor (Section 1.3). Our experiments use fixed public seeds so that the pools are reproducible.

A cell released unchanged does give the attacker the original text, but not the randomness used to generate the owner's candidates. The attacker can generate new paraphrases with the same public model. That is the fresh-paraphrase attack we evaluate: three rounds at the owner's temperature of 1.2 leave the score above 3.09 in Figure 4, and a different paraphrasing model leaves z = 20.3 on Airbnb (Table 17). Reselection from the exact owner pool, which can erase the mark, grants access beyond the threat model; we keep it as an upper bound on attacker knowledge.

Q3 (controls and fixed keys). Tables 16 and 17 report 24 distinct clean and wrong-key controls, with no flags. They are not independent trials, since the wrong-key tests reuse watermarked releases, and even if all 24 were independent, zero flags would give only an 11.7% one-sided 95% Clopper–Pearson upper bound. The false-positive rate rests on calibration, described next.

For text, the guarantee is over the random draw of the owner's key. We score the suspect column under 20,000 fresh keys and form its empirical tail p-value, including the owner's key as one of the exchangeable draws under the null. For a table and descriptor independent of the secret key, this gives a valid test over the key draw. A guarantee over innocent tables for each fixed deployed key is not established. Section 1.3 states this scope. Appendix D.1 additionally reports 10,000 independent clean draws with no text score above 3.09.

For numbers, a fixed deployed key needs calibration on held-out clean tables from the same distribution. The random-key normal-tail threshold does not by itself give that guarantee: across 50 fixed keys, each tested on 2,000 clean tables, the observed false-positive rate at z > 3.09 ranges from 0% to 1.20% (Table 11). Per-key calibration takes about four seconds for 2,000 tables. We reran every mixed-table decision in Tables 16 and 17 with thresholds calibrated on 20,000 held-out clean tables per deployed key, and no verdict changed. The smallest control p-value was 0.017, above the Bonferroni threshold of 3.3·10⁻⁴ (Appendix F.1.1). This calibration assumes representative clean data.

Thanks for the rebuttal effort and clarification. After reading the response and other reviewers' comments, I think the main text writing and presentation still needs more effort. Therefore, I will keep my score.

Reviewer_dyypOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces HashMark, a watermarking method for tables containing numerical values and free-form text. It makes small numerical edits and selects text paraphrases using a secret key. Detection requires the suspect table, the key, and a public descriptor, without access to the original table or its generator.

Critical strengths
  1. Watermarking tables containing both numerical values and free-form text is a useful practical problem.

  2. The proposed watermark method does not need the original table or the tabular generator for detection.

Critical weaknesses
  1. The writing of the paper needs significant improvement. For example, the main text does not explain technical concepts where readers need them. In Theorem 1, what is the random-function model? In Section 2.3.2, the sentence: 'Under uniform sampling, the valid fraction stays the same in expectation while the score falls with the square root of the number of items left; this describes the mean, not every subset' is really confusing to read. In Proposition 3, what is Hypergeom?

  2. My other major concern is about experiments: the experimental presentation is fragmented, making the claims difficult to assess. Section 3 moves between numerical datasets in Table 2, text corpora in Figure 3, mixed tables in Table 3, and separate attack studies. Experiments also use different method variants and detection criteria, with essential details scattered across the appendices. A consolidated summary should identify the dataset, method variant, attacker access, attack strength, and detection rule for each main experiment.

  3. The method cannot be applied to tables with purely categorical columns, which limits its practical usage.

Questions for authors

See weakness.

3 comments

We thank the reviewer for the careful review. We hope the replies and revisions below resolve their concerns and merit a reassessment of the score.

W1 (writing). The examples you identify should have been explained where they occur. We have revised each passage and added definitions at first use elsewhere in the paper.

In Section 1.3 and Theorem 1, we now state the random-function assumption explicitly: over a random key S, the keyed hash H(S, ·) assigns independent uniform bits to distinct inputs. Each previously unqueried grid value therefore gives the numerical search a fresh chance of acceptance. Our implementation uses HMAC-SHA-256 as a pseudorandom function; Appendix C.1 explains the corresponding computational qualification.

We have replaced the survival sentence in Section 2.3.2 with: "If an attacker keeps a uniformly random subset of the items, the expected fraction of valid items does not change. The score is the deviation of that fraction from one half, scaled by the square root of the number of items, so it decreases as the square root of the number of items kept." The next sentence states that this describes the average subset, and Proposition 3 gives the full distribution.

Immediately before Proposition 3, we now define Hypergeom(m, T, m'): it is the distribution of the number of valid items obtained when sampling m' items without replacement from m items, T of which are valid.

W2 (fragmented experiments). We have added the consolidated summary you requested as Table 2 in Section 3. It lists the dataset, method variant, attack setting and strength, detection rule, and metric for each experiment. This includes Edit Mode, HashMark-Q, text-only and mixed-table experiments, and the composed baseline. Each experiment has an ID linking it to its detailed results; Section 3 uses the same IDs when discussing the evidence. Table 5 in Appendix A provides a corresponding map from claims to supporting experiments.

We also distinguish three uses of scores: calibrated table-level ownership decisions, TPR at the benchmark's 1% FPR operating point, and descriptive raw z-scores. The experiment table states which applies, so a descriptive score is no longer presented alongside a detection rate without explaining the difference.

W3 (categorical columns). HashMark can watermark the numerical and free-form text columns of a table that also contains categorical columns. The categorical columns are released unchanged and contribute no evidence to detection. All our mixed tables include categorical columns (Table 4; Appendix B.3, Table 6). A purely categorical table cannot carry a HashMark watermark; we agree that this limits its scope and have stated it explicitly.

Encoding categories as integers does not resolve this limitation without a cost. TabularMark suggests such an encoding but does not evaluate it. Changing an integer code changes the category, and nearby codes need not have similar meanings. Adding fractional digits preserves the recoverable label but creates an obvious pattern that rounding removes. Appendix B.3 explains these two cases. We therefore leave categorical labels intact.

Thanks for your response.

For W1, the modification helps clarify the concepts I mentioned, but I still believe the overall readability is quite limited. I encourage the authors to revise/reorganize the paper thoroughly and submit it to future conferences.

For W2, the provided summary does not make the experiments less fragmented; it only lists all the experiments set up in a plain way; my concern is not addressed.

For W3, as agreed by the authors, the proposed method cannot be applied to a table with pure categorical columns, which is a big limitation for practical usage.

Based on the above reasons, I will maintain my score.

Thank you for the reply. We respectfully disagree on all three points.

W1. Your review named three passages: the random-function model in Theorem 1, the survival sentence in Section 2.3.2, and Hypergeom in Proposition 3. You agree that the revision clarifies them. We also defined every technical term at first use. We are happy to make further edits based on suggestions, within the discussion period.

W2. Your review asked that "a consolidated summary should identify the dataset, method variant, attacker access, attack strength, and detection rule for each main experiment." Table 2 does exactly this, one row per experiment, and Section 3 is built around it. The section answers three questions in order: how much of a table can be marked, what survives editing, and how HashMark compares with existing watermarks. Every result cites the experiment ID that Table 2 defines.

The experiments use different variants because each answers a different question:

  • Edit Mode for exact fidelity
  • HashMark-Q for noise
  • The text channel for paraphrasing
  • Mixed tables for the joint decision

Each detection rule is the one its comparison requires, for example the 1%-FPR operating point of the MUSE benchmark. We do not see how presenting these results under one variant or one rule would make them easier to assess. If you have a specific organization in mind, we will adopt it.

W3. A purely categorical table holds its information only in its labels, so any watermark for it must change labels. Changing labels runs into two obstacles that together leave little room for a useful watermark.

First, a changed label is a false record: a wrong diagnosis, a wrong marital status, a wrong product category. HashMark is designed so that no released record is false. Numbers change only below the declared precision, and text changes only within its meaning. Section 1.1 therefore leaves categories exact.

Second, a label cannot carry a mark by itself. A numerical or text cell has many possible values, so its hash alone carries evidence, and it survives the deletion of any other column. A label has only a few values. If the key fixed which labels are valid, marking would move rows into the valid labels and visibly shift the column's distribution. Validity must therefore depend on the label together with other columns of the row. The mark then lives in the relation between columns, and an attacker who deletes a context column deletes the mark.

We measured both obstacles. On the categorical columns alone of Shoppers (seven) and Adult (eight), the key selects about γ of the rows (2%, 5% or 10%) from a hash of the row's other labels, and a second hash marks each label valid or invalid for that row. A selected row with an invalid label receives the most frequent valid one. Detection compares the count of valid labels among selected rows with the counts under 200 fresh keys.

The mark is detected and the aggregates barely move. Changing 74 Shoppers TrafficType labels (0.9% of rows) gives z = 12.2 against a fresh-key standard deviation of 2.95, and changing 523 Adult occupation labels (2.5%) gives z = 32.1 against 5.9; no fresh key reaches either. Across twelve settings changing 0.7% to 5.5% of rows, downstream accuracy moves by at most 0.34 points over five splits, and Cramér's V between the marked column and any other changes by at most 0.029 (Appendix B.3).

The mark does not survive the deletion of a column. Dropping any one of the six other Shoppers columns raises the fresh-key p-value from 0.005 to between 0.08 and 0.82. Dropping any one of the seven other Adult columns raises it to between 0.05 and 1.0. Keying on a single context column, which deletion of the other columns cannot affect, makes the mark visible. On Shoppers, with Region as context, the key selects whole regions at once, so marking changes 134 labels and increases the shift in Cramér's V eightfold, to 0.014; dropping Region then gives p = 0.99. On Adult, with education as context, no education level falls below the selection rate, so no row is marked.

Each of those 74 Shoppers sessions and 523 Adult respondents now carries a label the owner knows to be false, and a single deleted column removes the evidence that justified it. A categorical watermark therefore falsifies records in exchange for a mark that is either visible or fragile. We do not consider that a good trade, and we leave categories exact. On Wine, for the same reason, we also leave whole-dollar price exact, since marking it by unit steps leaves a visible pattern (Appendix F.1), and the description alone carries the mark.

Your review identified tables with numbers and free-form text as a useful practical problem. This paper solves that problem, and no prior method marks such tables under one key with one decision. Leaving categories exact follows from the two obstacles above and does not, in our view, warrant a clear rejection.

We ask you to reconsider your score in light of these points.

Comments on the paper

1 comment

We thank the reviewers for their comments. We revised the paper to make the experimental evidence easier to follow, clarify what the statistical guarantees cover, and address fidelity and robustness concerns more directly. The revised main text uses the additional page allowed for the revision; section, table and appendix numbers below refer to the revised PDF.

The largest presentation change is a consolidated experiment table in Section 3 (Table 2). It identifies the dataset, method variant, attack setting, detection rule and metric for each experiment, with references to the detailed results. We also define the technical terms at first use and separate calibrated ownership decisions from descriptive scores. These changes address dyyp's concerns about the writing and experimental organization, as well as Kfei's concern about the different metrics.

For Kfei's questions about false positives, we now distinguish guarantees over random keys from calibration for a fixed deployed key. We reran all mixed-table decisions with per-key numerical calibration; no verdict changed (Appendix F.1.1). We also clarify that the text test is calibrated over the key draw and does not establish a false-positive guarantee over innocent tables for every fixed key. Section 1.3 now states this distinction and explains how the owner keeps paraphrase-generation randomness secret while using public models.

For cWkr's concerns about robustness and semantic fidelity, the revision adds operating points showing the numerical fidelity–robustness trade-off, a blinded human evaluation of rewritten cells, and medical-text experiments with candidates generated under explicit factual-preservation instructions. The results support using better candidate generators, but they also show why factual preservation cannot be assumed from embedding similarity alone. Our reply to cWkr discusses these findings and their limitations.

The changes and their locations:

  • Presentation and experiment maps (dyyp W1, W2; Kfei). Table 2 (Section 3) indexes every experiment by ID; Table 5 (Appendix A) maps claims to appendix evidence with the same IDs. Definitions at first use: the independent-bit assumption (Section 1.3, Theorem 1), Hypergeom and the survival paragraph (Section 2.3.2), the family-wise decision (Section 2.2), C2ST, MLE gap and null standard deviations (Section 3). Table 4 labels each p-value; descriptive z-scores move to Table 14.
  • Guarantees and calibration (Kfei questions 2 and 3). Secret paraphrase randomness, per-key numerical calibration and scope of the text guarantee (Section 1.3, Appendix D.1.1, Table 11); 10,000 clean text draws (Appendix D.1); control count (abstract, Table 4).
  • Positioning and scope (cWkr W1, dyyp W3). Table 1 adds WGTD, SemStamp and a cross-column-type row; Section 1.2 compares with MUSE and Ngo et al.; Appendix J.2 reimplements Ngo et al.; Appendix F.1.2 reports the composed baseline; Appendix B.3 states the categorical scope; Appendix F.1 marks an integer column.
  • Fidelity and robustness experiments (cWkr W2, W3). Appendix K.1 adds the entailment-filtered release, the blinded human evaluation, and the medical-text experiment with LLM-written candidates (Table 40); Table 44 (Appendix L.1) gives operating points by bin count; Appendix E.2 adds the few-paraphrase capacity estimate.
  • Corrections. The 0.38-point utility figure is the bound of the 95% intervals on three UCI tables; the 32-bin noise claim holds on Shoppers; the 0.86-point Wine cost is at 30.5% of descriptions rewritten; the Figure 4 reference line is z = 3.09.
  • Code. The anonymous repository linked in the Reproducibility Statement now includes the code for every new experiment: the Ngo et al. reimplementation, fixed-key and per-key calibration, the filtered release, and the medical-text and LLM experiments.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.