acceptodds
Under review as a conference paper at ICLR 2027

Tabular Pancakes: Cryptographic Watermarks Against Dataset Fabrication

Abstract

Tabular generative models can fabricate and impute scientific datasets, bypassing fraud forensics and heightening malpractice risks. Since no statistical test on outputs alone can expose a well-trained generator, provenance must be embedded at generation time. We introduce Tabular Pancakes, a training-free watermark that replaces the generator's Gaussian latent prior with the homogeneous CLWE distribution: Gaussian in all directions except a secret subspace, where it is periodic. With the key, an auditor efficiently recovers the watermark from a small fraction of released rows, and every verdict carries a certified p-value. Recovery is invariant to row permutation, column renaming, and rescaling. Under tampering such as cell edits or re-imputation, detection power decreases only linearly in the fraction modified—no small edit removes the mark. Row-level tests localize synthetic rows in mixed tables with calibrated false-discovery rates, turning detection into forensic evidence. The mark can equivalently be embedded directly in released rows, covering generators whose samplers cannot be inverted. Without the key, marked and unmarked outputs are indistinguishable, and keyless removal provably requires √n/k-fold larger distortion than keyed removal, ruling out selective scrubbing. We verify Tabular Pancakes on four datasets and six generator families (TabSyn, TabDDPM, STaSy, CoDi, SMOTE, GReaT), towards making data fabrication provably traceable.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
watermarking, generative tabular models, hCLWE

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_fFbEOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
3
Presentation
2
Soundness
3(+1)
Read review
Paper summary

The tabular generative models in use nowadays are able to fabricate or impute realistic scientific datasets effortlessly. These datasets can bypass traditional statistical fraud forensics i.e., Benford’s law and GRIM or SPRITE checks, creating an urgent need to embed verifiable provenance at generation time rather than relying on post-hoc detection. However, existing tabular watermarks (i) can be detected and scrubbed by adversaries, (ii) lack the rigorous false-positive guarantees needed for research misconduct investigations, (iii) and break under routine table formatting edits. To solve this, the authors introduce TABULAR PANCAKES, a training-free cryptographic watermark that replaces the generator's standard Gaussian latent prior with a lattice-based distribution that looks like ordinary Gaussian noise in all directions except a secret periodic subspace, alongside a direct data-space variant for non-invertible diffusion models whose reverse trajectories severely amplify inversion errors. Evaluations across four datasets and six generator families show that the watermark (i) is cryptographically undetectable without the secret key, (ii) provably requires substantially higher distortion to scrub without the key than with it, (iii) remains strictly invariant to row shuffling, column renaming, and unit rescaling, and (iv) provides certified statistical significance alongside row-level localization to pinpoint synthetic records inside partially fabricated tables.

Critical strengths

The motivation behind the study is well laid out and proves why a study in this direction is necessary. The method proposed by the authors is built upon cryptographic and statistical foundation tailored for forensic auditing. Basically, rather than relying on heuristic markers that classifiers can easily detect and scrub, the authors ground their scheme in lattice hardness which provides evidentiary standards that include certified false-positive control and robust invariance to routine table edits. Also, another vital technical contribution by the authors is the identification of the latent inversion barrier in diffusion-VAE pipelines which proves that reverse ODE trajectories amplify errors by up to 100x and wipes out fine phase signals. Based on these contributions, I believe the authors have shown the significance of thier study.

Critical weaknesses

The fundamental issues of watermarking still exists. (1) Retraining a model using only the marked synthetic data destroys the mark, forcing the authors to rely on lower-bounding the high utility and distortion costs of such removal rather than preventing it entirely. (2) Datasets with very few continuous or numeric columns heavily constrain performance, they have mentioned this in their limitations, but this shows that their method is only useful for continuous dataset. Furthermore, they didn't evaluate on deterministic datasets (i.e., datasets with columns which are generated based on other columns, such as a column which is linearly dependent on the combination of other columns). These kinds of datasets may be continuos in nature, but what happens when their approach is used for marking such?

Questions for authors

Please check weaknesses above.

4 comments
AuthorsOpenReview ↗

Author Response (1/2)

Thank you for your positive and insightful review! We are encouraged that you found the study well-motivated and our lattice-based construction a sound foundation for forensic auditing. In response to the two weaknesses, we conducted new experiments on (i) mark survival under student-generator retraining and (ii) marking of dependency-constrained and mixed-type tables. They quantify the cost of removal (W1) and extend the codec to preserve exact column dependencies (W2). We summarise results below and revise Sections 3.4, 5.4, 6 and the Appendix accordingly:

W1: "Retraining on the marked data destroys the mark."
To summarise, we find that retraining does not reliably destroy the mark, and when it does, this does not come for free. Students that preserve data quality inherit a certifiable mark; students that shed it measurably degrade the data. To test this, we trained five student generators on marked output (3 datasets x 3 train sizes 6k-60k x 3 keys x 3 student seeds; 27 certified runs per cell) and ran the deployed certificate on each student's own samples (adult shown; full grid in App. D).

StudentResidual SSmedian ppcopy-sharecol-shape Δ\Delta vs teacher
CTGAN0.0050.430.00-0.26
TVAE0.0510.280.01-0.29
Copula0.0100.170.01-0.00*
SMOTE0.2690.0080.037-0.01
ARF0.3820.0020.005-0.07

*The Gaussian copula reproduces each column's marginal exactly by construction, so the column-shape metric cannot register its cost; the cost appears in cross-column structure, as a 0.05 drop in downstream prediction accuracy. Concretely:

  • 1.1 Removal is not free. The students that strip the mark pay a measurable fidelity price (CTGAN/TVAE column-shape -0.26/-0.29; copula downstream -0.05), consistent with our lower-bound analysis: scrubbing by retraining is a quality vs. provenance trade. Yet more interestingly,
  • 1.2 Removal is not guaranteed. The utility-preserving students inherit a certifiable mark. SMOTE certifies on all three datasets, and on adult only 3.7% of its rows are near-copies of training rows: the certificate is not firing on leaked memorised rows, but on rows the student synthesised itself, i.e. it has learned the marked distribution and reproduces the periodic structure in new samples. ARF, a bona-fide tree-ensemble generator, certifies at S=0.382, p=0.002 with copy-share 0.005; this indicates radioactivity—in the sense of Sablayrolles et al. (2020)—through genuine retraining.

We include and discuss these results in Sec. 5.4. The watermark is a component of the generator's output distribution: a low-amplitude periodic modulation of the density along the secret directions. A well-trained student can match that distribution and reproduce the modulation; therefore, the certificate fires on the student's own samples (SMOTE, ARF). If it cannot represent the modulation, its output distribution misses a component of the training distribution, and the gap surfaces as the fidelity costs in the table (CTGAN, TVAE, copula). Shedding the mark by retraining would require an imperfect fit of the training distribution and triggers a removal--fidelity trade-off. This trade-off is governed by the mark’s amplitude (γ\gamma, β\beta, kk), which is chosen by the auditor. While retraining cannot be wholly prevented, our results show that under Tabular Pancakes, this would result in either a certifiable residue or a measurable quality gap. We note that the false-positive guarantee holds unchanged in both because it rests on key rerandomization alone and does not rely on assumptions about the adversary.

Removal without retraining is also now measured in full. Gradient search with up to 10^6 restarts fails to locate the latent secret subspace (zero recoveries; ascent converges only from starts within ~0.2 rad of the true direction). Across seven budgeted removal attacks, even direction-informed attackers need cell-level RMSE of 0.05-0.30 sigma to defeat the certificate, and model-based re-imputation never succeeds within budget, for the same reason SMOTE and ARF inherit the mark above: the imputation model re-learns the marked conditionals. The complete results are in App. D.

[1] Sablayrolles, A., Douze, M., Schmid, C., & Jégou, H. (2020, November). Radioactive data: tracing through training. In International Conference on Machine Learning (pp. 8326-8335). PMLR.

(1/2; dependency-constrained and categorical data in the next reply)

AuthorsOpenReview ↗

Author Response (2/2)

W2a: "Only useful for continuous datasets."
We agree that the method is best suited to tables with mostly numeric columns and state this in the revision. Categorical columns, however, are not unused: they carry the keyed synchronisation signal, and we add a new experiment characterising when this channel remains undetectable.

We investigate collision as a possible failure mode: rows that induce the same synchronisation pattern could be grouped and compared, and systematic agreement within groups could reveal the mark. We measure this with category-conditional two-sample tests. Here kk denotes the number of secret marking directions, i.e. the dimension of the hidden periodic subspace; the synchronisation pattern takes one value per direction, so the number of distinct patterns grows exponentially in k and repeated patterns become rare. We include k=1k=1 as a theoretical worst-case baseline; here, the channel indeed leaks and the marked table contains 2.77M colliding row pairs. Notably, however, the leak decays with the collision count: at k=8k=8 the tests are at chance (AUC 0.50-0.51; 5 colliding pairs), which is in agreement with the analytically predicted counts. With a simple mitigation technique---additional synchronisation levels and binning---using k≥4k \geq 4 already suffices while using k≥8k \geq 8 provides strict indistinguishability. The revision therefore prescribes k≥4k \geq 4 with this mitigation for mixed-type tables, which coincides with the configuration recommended for localization and therefore, adds no new constraint in practice. Because the collision count is computable from the key parameters and the table schema, the safety condition can be verified before deployment.

W2b: "What happens on deterministic or linearly dependent columns?"
To answer this question, we constructed a suite of tables with exact column dependencies (linear combinations, sums, integer relations, and redundant columns with small jitter). The default codec is designed for unconstrained tables and is not the appropriate tool here; for dependency-constrained data the revision adopts a rank-restricted codec, which preserves the constraints by construction. An exact linear dependency among columns corresponds to a zero-variance direction of the data covariance. The rank codec performs PCA whitening truncated at a threshold and embeds the mark only in the retained subspace. The constraint null space has zero variance and is therefore never touched: the watermark cannot alter any direction in which the data does not vary, and dependent columns are reconstructed exactly from the retained components. Evaluation on the suite confirms the guarantee:

Suite with exact dependenciesrank codec (=1e-3)
constraint violations0.000 (pre-formatting drift < 1e-10)
detection statistic S0.846 (vs. 0.867 unconstrained; cost 0.021)
redundant-column jitterS = 0.861 (default codec: 0.645)

Two refinements remain and are left for future work. First, the final formatting step (clipping to the observed data range) perturbs 3-4% of rows, leaving a secondary statistical trace; constraint-aware formatting would remove this step's effect. Second, columns supported on coarse integer grids weaken detection (S=0.256S = 0.256); this is the same integer-quantisation regime as in W2a and motivates integer-lattice codecs, a direction we consider highly promising given the method's lattice foundation.

Summary.
We present three supplementary experiments. Our retraining study shows that removal by distillation is either paid for in fidelity or leaves a certifiable residue. The dependency suite introduces the rank codec, which preserves exact constraints by construction at negligible detection cost. The collision analysis equips the categorical channel with a safety condition, k≥4k \geq 4, that can be verified before deployment. Together, these results extend the paper's certified provenance guarantees (keyed undetectability, exact false-positive control, invariance to routine edits) to structured and mixed-type data. They also sharpen the comparison with latent-space schemes: under the deployed decode-and-invert round trip, no latent-space watermark certifies at any m≤16,000m \leq 16,000 rows, whereas the data-space mark certifies from m=125m=125 and is unaffected by shuffling, deletion, cell noise, and rescaling.

We hope these results provide a rigorous basis for data provenance in scientific forensics. Thank you for your insights; we hope these revisions helped address your concerns and welcome further discussion!

(2/2)

Nice! even though it came close to the end of the discussion window. Thanks for the clarification with new experiments, and I wasn't expecting the algorithmic extension. The new experiments directly address my concerns. Please do well to add them to the revision. I will update my soundness and overall rating to further support acceptance.

Thank you for actively participating and updating your score! We are delighted that our rebuttal addressed your concerns and will certainly incorporate these changes in the revision.

Reviewer_5k3COpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces TABULAR PANCAKES, a watermarking scheme for synthetic tabular data that replaces a generative model's Gaussian latent prior with an hCLWE (homogeneous Continuous Learning With Errors) prior — standard normal in all but k secret directions, along which the density concentrates on equally spaced Gaussian "pancakes." Because distinguishing hCLWE from N(0, I_n) is as hard as worst-case lattice problems (Bruna et al., 2021; Gupte et al., 2022), a keyless adversary sees only Gaussian noise: the mark is cryptographically undetectable, not merely empirically so.

The contributions are:

Embedding (Sec. 3): a training-free prior swap requiring no retraining or architecture change, exactly invariant to row order and per-column affine maps.

Security and detection theory (Sec. 4): undetectability reduces to lattice hardness (Thm. 4.1); false accusations are controlled with no distributional assumptions (Thm. 4.2(ii), certified null over key randomness only); detection, localization, and payload admit closed forms; keyless removal provably costs sqrt(n/k)-fold more distortion than keyed removal (Thm. 4.4); gamma controls the security–robustness tradeoff.

A conditioning barrier (Sec. 3.3): on diffusion–VAE pipelines the release map is not invertible at phase precision — the reverse trajectory amplifies input error by 50–100x independent of step count — so no verifier can recover the latent. The paper proves this (not just observes it) and responds with a data-space variant (Sec. 3.4) that embeds the pancake structure in a whitened representation of the released table, preserving first and second moments exactly, plus a categorical channel using keyed Gumbel-max.

Validation (Sec. 5): four TabSyn pipelines, six generator families, false-positive audits, adversarial learnability (C2ST at chance), and localization with FDR control and unbiased fraction estimation.

The paper is dense but well organized, the theory is unusually careful (full proofs in App. A, hardness statements in App. B, concrete-security analysis in App. F), and the negative result in Sec. 3.3 is itself a contribution the community should know about.

Critical strengths

S1. hCLWE is a prior that inherits a lattice hardness reduction rather than a heuristic, which is what a forensic mark has to be. The conclusion puts the inversion better than I could: the fidelity that defeats statistical forensics is what certifies provenance.

S2. 4.2(ii) is the one that matters in practice. The null holds over the verifier's key randomness alone, with nothing assumed about the data, so false accusations sit at alpha whatever the suspect table holds. TabWak, Tree-Ring and Gaussian Shading have no equivalent.

S3. 3.3 is proved, not measured. sigma = 1.2–1.6 at 50 steps, sigma_ode = 0.78 at 50, 200 and 1000, which rules out discretization; the expanding map is the culprit. The sign-phase divide generalises: sign schemes need O(1) latent accuracy, phase needs O(1/gamma') and is unrecoverable.

Critical weaknesses

W1. "Cryptographically undetectable" in the title describes the latent variant; the variant behind Tables 2–3, and the only one that works on TabSyn, is the data-space one, whose undetectability is audited rather than proved. A sentence in the abstract and one in Sec. 1 naming which variant carries the claim would fix most of it. The audit also undercuts the rhetoric as printed: logistic C2ST sits at chance (0.501 / 0.496 / 0.508 / 0.504) while HGB reaches 0.962 on Adult, 0.746 on Magic, 0.993 on Default and 0.920 on Shoppers, and those are the numbers an informed adversary cares about. Either lead with them and argue the threat model, or write "undetectable against a keyless adversary".

W2. The deployed parameters sit far outside the regime the theorems describe: gamma >= 2*sqrt(n), about 16 at n = 64, against gamma = 2 deployed; k = Omega(log N) against k <= 4 at m about 6000. Remarks A.2 and 4.3 say this honestly, the main text does not. State once whether gamma = 2 is undetectable or undetectable against the attacks tried, and note that the concrete analysis rests on the covariance attack alone.

W3. The TabWak comparison is asserted, not measured. That its sign-based marks are separable by a classifier is a strong claim about the closest prior work, and your own experiments fork their pipeline (App. G). A head-to-head C2ST with AUC on TabWak-marked against unmarked rows would cost little and carry the claim.

Questions for authors

On W1/W2 (which variant does the headline claim cover): Can you state, in one sentence in the abstract, that the cryptographic undetectability guarantee applies to the latent variant (used when the release map is invertible) and that the data-space variant (used on diffusion–VAE pipelines) is audited empirically? And can you report the HGB C2ST numbers alongside the logistic-regression numbers in the main text, with an explicit statement of what the keyless adversary knows (in particular, whether the marked column list is part of the key and hence hidden)?

1 comment
AuthorsOpenReview ↗

Author Response

Thank you for this careful and insightful review. We are delighted that you found the hCLWE foundation, the assumption-free certified null, and the conditioning-barrier proof to be critical strengths of our paper. Based on your suggestions, we clarify the scoping of the undetectability claim, the stated security regime, and the substantiation of the TabWak comparison; we revise the paper to adopt the three proposed fixes.

W1 (which variant carries the headline claim).
We agree: the cryptographic claim belongs to the latent variant, and the revision states this clearly. The abstract now reads "Without the key, marked and unmarked outputs are computationally indistinguishable for the latent variant and audited empirically for the data-space variant"; Section 1 attributes each guarantee to its variant; and a new table in Section 3.4 lists, per variant, what is proven and what is audited. Section 5.4 reports the HGB numbers next to the logistic-regression ones, with the threat model stated explicitly: the adversary knows the generator, schema, codec eligibility rule, and all public parameters, while the key K=(W,γ,β,κ,δ,scope)K = (W, \gamma, \beta, \kappa, \delta, \mathrm{scope}) is hidden, where the scope descriptors include the marked column list; the marked column list is part of the key. To characterise what the HGB signal contains, we additionally ran matched-distortion controls (Appendix D): RMSE-matched unkeyed noise is detected within 0.03-0.05 AUC of the real mark on every dataset. An informed adversary running HGB therefore learns that the table was perturbed, which any watermark at equal distortion reveals, and gains at most 0.05 AUC from the keyed structure itself. Where the data-space variant is discussed, we adopt your wording of "undetectable against a keyless adversary".

W2 (deployed parameters vs. the asymptotic regime).
We revise the main text to state explicitly what was previously noted only in Remarks 4.3 and A.2: at the deployed point (γ=2\gamma = 2, k≤4k \le 4), undetectability is a concrete-security claim, i.e. secure against the attack landscape tested under per-key release caps, while the hardness reduction governs the asymptotic regime 2n. The tested landscape is also broader in the revision and no longer rests on the covariance attack alone: a new suite (Appendix D) adds gradient search on the detection score with up to 106 restarts (zero recoveries), a surrogate distinguisher trained on up to 105 attacker samples (AUC 0.50 throughout), and a seven-attack removal frontier that includes direction-informed attackers. Appendix F now cites this suite as the evidence behind the concrete claim.

W3 (TabWak comparison).
We conduct a head-to-head comparison with TabWak, and observe that Tabular Pancakes retains certified verification under deployment where TabWak does not. TabWak's mark is only weakly separable at the released-table level, but it is exposed once an attacker gains latent access and does not certify through the deployed verification path, whereas the data-space mark does. Concretely, through the shared verifier on TabSyn-Adult (Table 1), TabWak-marked vs. unmarked rows give C2ST 0.511 (logistic) and 0.540 (HGB); the revision extends this to Magic and Default with the same conclusion; we reword our related-work claim to match the measurement. With inversion access, an attacker separates TabWak at AUC up to 0.89 and Gaussian Shading at 0.92-1.0 (Appendix E). On the verification side, through the decode-invert round trip no latent scheme, TabWak included, reaches 95% detection at any m≤16,000m \leq 16,000 rows, whereas the data-space mark certifies from m=125m=125 (Appendix E). We discuss the underlying cause of both effects in Section 3.3: phase information does not survive the release map.

Re: question on the abstract sentence and the HGB numbers. Yes to all parts. The abstract now states, in one sentence, that computational indistinguishability applies to the latent variant (applicable when the release map is invertible) and that the data-space variant (deployed on diffusion-VAE pipelines) is audited empirically. Section 5.4 reports the HGB C2ST numbers alongside the logistic-regression ones, together with the threat-model statement above: the keyless adversary knows the generator, schema, codec rule, and all public parameters; the marked column list is part of the key and hence hidden.

We hope the revision resolves the scoping concerns. We thank you again for a review that significantly improved the paper's precision and would welcome further discussion!

Reviewer_8V9wOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper embeds keyed periodic structure in Gaussian latents and proposes a data-space fallback when latent inversion fails. It develops detection, localization, and removal-cost claims, with controlled-flow experiments and evaluations across tabular generators.

Critical strengths
  • The analysis of phase-recovery failure in TabSyn is informative. The exact key-randomization test and mixed-table localization provide useful forensic components beyond a single aggregate detection score (Sections 3.3 and 5.3).
Critical weaknesses
  1. The deployed variant lacks the central cryptographic guarantee. Latent verification fails on TabSyn, while data-space embedding has no hardness reduction. HGB C2ST scores of 0.746-0.993 in Table 2 contradict Section 5.4's indistinguishability claim. The conclusion overstates which guarantees transfer to real generators.
  2. Theorem 4.2 states an exp(-mkS^2/2) certificate for arbitrary key-independent data, but Appendix A.2 requires additional product-regime assumptions. Repeating an identical row does not create independent evidence under a shared key. The unconditional claim and sample-complexity statement therefore exceed the proof; the separate Monte Carlo certificate can remain valid.
Questions for authors
  1. Is there a deployed operating point that simultaneously gives useful localization and near-chance performance for the strongest tested distinguisher?
  2. Can Theorem 4.2 separate exact randomization guarantees from concentration bounds under explicitly justified independence and bias assumptions?
3 comments
AuthorsOpenReview ↗

Author Response (1/3)

Thank you for your constructive feedback. We are glad you found the conditioning-barrier analysis and the forensic components useful—Tabular Pancakes is to our knowledge the first tabular watermark whose verdicts are certified p-values with row-level attribution rather than a single aggregate detection score. Your comments have helped us tighten the statement of Theorem 4.2 and clarify the scope of the undetectability claims. To verify them empirically, we add three experiment suites: a null-calibration stress study (which includes the duplicated-row case you raise), a detectability study with matched-distortion controls, and an attack suite (Appendices D and E). Below we address W1 and Q1; W2 and Q2 follow in the next reply.

W1: "The deployed variant lacks the central cryptographic guarantee; Table 2's HGB scores contradict Section 5.4."

  • (1.1) Which guarantees transfer, stated per variant. The revision adds a per-variant guarantee table (Section 3.4). The guarantees that carry the forensic verdicts transfer to the data-space variant unconditionally: exact false-positive control and unforgeability hold for any suspect table, because the deployed verdict is the key-rerandomization Monte Carlo certificate. This is exact even without distributional assumptions (reply 2). Undetectability for this variant is audited empirically, as specified in Section 3.4. The revision aligns Section 5.4 with this audit: the indistinguishability claim is now stated per distinguisher class, holding at chance for the logistic-regression C2ST of Table 2, while the signal seen by stronger learners is analysed in the next item, where matched-distortion controls show that at most 0.05 AUC of it is attributable to the key.

  • (1.2) What the detectability consists of, and what it buys (new controls and a laundering study). We ran matched-distortion controls on the three TabSyn datasets: V5/V6 replace the mark with RMSE-matched unkeyed noise (isotropic, or along fresh random directions). Unkeyed noise is detected within 0.03-0.05 AUC of the real mark on every dataset, including the integer-heavy ones: the strongest classifiers are detecting that the table was perturbed at all, a property shared by any watermark at equal distortion, and the advantage attributable to the keyed structure is at most 0.05 AUC everywhere. The revision reports detectability as this decomposition (perturbation + keyed residue), which we believe is the informative quantity for a keyed scheme. The same learnability has a forensic benefit: a signal that a classifier can pick up from the released table is also one that a generator trained on that table picks up and reproduces. In a new study across five student generators trained on marked output (27 certified runs per pair), the utility-preserving students inherit a certifiable mark (SMOTE, adversarial random forests), while students that shed it degrade fidelity measurably (CTGAN/TVAE column-shape -0.26/-0.29). A student can only inherit the mark to the extent that the mark is learnable, so detectability and robustness to retraining are controlled by the same parameters (γ,β,k\gamma, \beta, k); the validity of the certificate in (1.1) does not depend on where this trade-off is set.

  • (1.3) Keyed secrecy, now verified end to end. The new attack suite validates the hardness assumption for the latent variant: an attacker holding everything but the key cannot locate the secret subspace. Gradient search on the detection score with 10^6 restarts recovers no direction, and the measured basin of attraction (~0.2 rad) is the empirical signature of the flat score landscape the reduction predicts: outside a vanishing cone around the planted direction there is no gradient signal to follow. A distinguisher trained on 10^5 attacker samples likewise stays at AUC 0.50. For the data-space variant with d≤10d \le 10 columns, subspace secrecy is impossible for any scheme, as the search space is too small. Hence the relevant guarantee there is the cost of removal, which the attack suite quantifies: even an attacker who has recovered the directions must corrupt every cell by 5-30% of a column std to defeat the certificate. Moreover, model-based re-imputation never succeeds, since the imputer re-learns the marked conditionals. In short, the latent variant provides computational hiding of the subspace when the latent dimension is large, and the data-space variant provides a measured removal cost when it is not.

(1/3)

AuthorsOpenReview ↗

Author Response (2/3)

Q1: "Is there a deployed operating point with useful localization and near-chance performance for the strongest distinguisher?"

Yes. On invertible-flow generators the latent variant achieves both: the strongest classifier two-sample test sits at chance (0.502 ±\pm 0.002) while localization attains row-level AUC >= 0.999 with FDR <= 0.046 and power 0.73 at a synthetic fraction of only 2% (Section 5.1). This is predicted by our theory: when the release map is invertible, the verifier recovers each row's latent at phase precision, so undetectability is inherited from the hardness assumption while every row still carries a score for attribution. Practically, this covers the partial-fabrication setting that motivates the paper, e.g. flagging a few dozen imputed records inside a table of thousands with certified FDR, and a provider who needs both properties can simply choose an invertible sampler at release time. On non-invertible hosts, the data-space variant at (2.0, 0.05, k=4) on all eligible columns gives localization power 0.973 at FDR 0.04 (row-level AUC 0.997), with the strongest distinguisher at 0.76, of which at most 0.03-0.05 is attributable to the key (V5/V6 above). The remaining signal is perturbation detectability, which no watermark at equal distortion avoids; with respect to the keyed structure the distinguisher is near chance. This configuration also improves both quantities at once relative to marking fewer columns: using all 10 columns instead of 2 lowers the strongest distinguisher from 0.886 to 0.760 and raises localization power from 0 to 0.973. Appendix D reports the full (γ,β,k\gamma, \beta, k) characterisation with per-seed values.

W2: "Theorem 4.2's exp⁡(−mkS2/2)\exp(-mkS^2/2) certificate exceeds its proof; repeated rows are not independent evidence; the Monte Carlo certificate can remain valid."

The Monte Carlo certificate is indeed the deployed one: Section 3.2 states "on real data we report the latter," and every p-value in Tables 1-3 is pMCp_{MC}; all reported verdicts rest on a certificate that is exact without distributional assumptions. We agree that the theorem statement should attach the premises of the analytic rates explicitly; the revised Theorem 4.2 does so. We take up the duplicated-row example you raise and test it directly, analytically in (2.1) and empirically in (2.2).

  • (2.1) Revised statement. Theorem 4.2(ii)-(iii) now presents its three certificates with their premises attached, following the structure of Appendix A.2: the exact key-rerandomization certificate, valid for any fixed suspect table because exchangeability holds over the verifier's keys rather than over the data. Also, an assumption-free concentration bound at rate k, where, as you observe, repeated rows add no independent evidence. Finally, the sharp exp⁡(−mkS2/2)\exp(-mkS^2/2) rate and its sample-complexity along with the product-regime premises (i.i.d. rows, independent per-direction scores, small bias term) under which they are proven. Reported verdicts use the first certificate exclusively; the second and third characterise detection power and yield the row-count requirements of Theorem 4.2(iii).

  • (2.2) The deployed certificate under adversarial nulls (new study). We measured false-positive rates on 13 null families designed to break row-independence: your example of a single row repeated mm times (with and without jitter), low-rank repeats, heavy mixtures, integer lattices, and unmarked real tables from four datasets (540 cells, 100k keys per cell, both verifier paths). Type-I error stays at the nominal level throughout: worst pooled rejection rate 0.0525 at =0.05, and 0.0496 in a replication with fresh null keys (n=30,000). These families defeat any test that treats rows as independent evidence; a row repeated m times carries the evidence of one row. The certificate is unaffected because its validity rests on exchangeability over the verifier's keys, not on any property of the table, so no suspect-table structure can inflate the false-positive rate. Forensically this means that the probability of a false accusation is controlled even on adversarially structured data. The duplicated-row case also confirms the rate-k bound: duplication raises the row count but not the k independent per-direction phases the bound counts. Full tables are in Appendix D.

(2/3)

AuthorsOpenReview ↗

Author Reponse (3/3)

Q2: "Can Theorem 4.2 separate exact randomization guarantees from concentration bounds under explicitly justified independence and bias assumptions?"

Yes; the revised Theorem 4.2 makes this separation explicit. The exact randomization certificate carries the verdicts. Its validity uses no assumption on the data: the suspect table may be arbitrary, only exchangeability of the true key among rerandomized keys is needed. The concentration bounds carry the power and sample-size statements. Their premises (i.i.d. rows, independent per-direction scores, small bias) are now stated in the theorem, verified on the invertible flow where ground truth is available (Section 5.1), and stress-tested by the null families above. We believe this separation—exact tests for verdicts and conditional bounds for planning—is suitable for data provenance forensics; we are not aware of a tabular watermark whose deployed verdict is an exact, assumption-free test.

Summary. Our revised manuscript attaches explicit premises to each certificate in Theorem 4.2, scopes the undetectability claims per distinguisher class (Sections 5.4 and 6), and adds three experiment suites substantiating both. These changes further strengthen the paper's central contribution: a watermark whose forensic verdicts (detection, attribution, and row-level localization) are exact p-values; valid for arbitrary suspect tables; complemented by the conditioning-barrier analysis which you highlighted.

Thank you, again, for your thoughtful suggestions; we hope our additions have addressed the soundness concerns and welcome further discussion!

(3/3)

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.