acceptodds
Under review as a conference paper at ICLR 2027

KASS: -Anonymity-Aware Sequential Sampling for Synthetic Tabular Data

Abstract

Synthetic tabular data generation seeks to preserve key statistical properties of the original data while reducing disclosure risk, yet existing generators may reproduce rare or unique attribute combinations and violate hard structural constraints. We propose KASS, a -anonymity-aware sequential sampling framework for synthetic tabular data. Rather than enforcing -anonymity as a post-processing step, KASS embeds its protective logic directly into the sampling process by restricting candidate values to attribute combinations with sufficient empirical support, making rare and potentially identifying patterns structurally unreachable. To preserve cross-variable dependencies and data plausibility, KASS employs normalized mutual information (NMI) thresholding to relax conditioning variables when strict conditioning becomes too sparse, and supports multiple banned-set strategies for controlling the privacy-utility trade-off. We analyze KASS from a differential privacy perspective and establish its relationship with -diversity. Experiments on real-world datasets show that KASS achieves competitive utility and distributional fidelity, produces substantially fewer constraint violations than competing methods, and generates no unique original samples.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Synthetic tabular data, k-anonymity, sequential sampling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_UhBUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

KASS is a kk-anonymity-aware sequential sampling framework that does not require training of generative models. It gradually samples from the empirical conditional distributions of the original data in random column order, and removes candidate values with support numbers below kk at each step; When the current conditions prevent further sampling, it relaxes the conditions from weakly correlated variables based on the NMI with the current variable, thus balancing privacy constraints with data distribution fidelity.

Critical strengths

1.This paper introduces a kk-anonymity-aware sequential sampling framework that provides explicit privacy control through parameter kk and NMI threshold, which can structurally prevent the generation of synthetic records corresponding to rare or unique individuals in the original data.

2.This paper analyzes from the perspective of differential privacy and discusses its relationship with l-diversity, providing theoretical support.

Critical weaknesses

1.KASS relies on empirical support counting and paired NMI estimation, both of which may be unreliable when the original dataset is very small. In this case, kk-anonymity-aware sequential sampling may become impractical, leading to frequent conditional relaxation or restarts, and potentially reducing data utility. Furthermore, this article does not determine whether the experiential privacy benefits persist under these low data regimes (<100), as reliable support estimates are particularly challenging. Evaluating small datasets will help clarify the robustness and practical privacy utility trade-off of this approach.

2.Using paired NMIs to guide conditional relaxation may not be sufficient to capture higher-order dependencies between attributes. Paired NMI measures marginal pairwise associations, which may not be able to identify dependency relationships that arise solely through joint interactions between multiple variables. Therefore, the proposed heuristic method may remove a condition variable that is crucial for maintaining high-order structure. Although the ablation study compared NMI based sorting with random dropout, it did not establish the reliability of this heuristic method under complex high-order dependencies. Evaluating synthetic data with controlled higher-order interactions will help clarify this limitation.

Questions for authors

1.How robust is KASS in very small-data regimes, where empirical support counts and pairwise NMI estimates may be unreliable?

2.Pairwise NMI may fail to capture higher-order or context-dependent dependencies. Have you evaluated KASS on datasets with controlled higher-order interactions?

3.Beyond Authenticity and UiOiS, have you evaluated KASS against stronger privacy attacks such as membership inference or attribute inference?

4 comments

Evaluation on the small datasets with higher-order interactions (Q1, Q2)

We evaluate KASS on simulated datasets with small sample sizes (N=100 and 500N=100 \ \text{and} \ 500) and higher-order interactions between variables. To simulate these higher-order interactions, we consider a scenario where the data samples are partitioned into two imbalanced clusters at an 8:2 ratio, with each cluster exhibiting a different correlation structure. Out of eight variables (X1,…,X8X_1,\ldots,X_8), variables X1,X2,X3,X4,and X5X_1,X_2,X_3,X_4, \text{and } X_5 are correlated within Cluster 1 (the majority cluster). On the other hand, within Cluster 2 (the minority cluster), variables X1,X2,X6,X7,and X8X_1,X_2,X_6,X_7, \text{and } X_8 are correlated. For the detailed data generation process, please refer to the earlier section, Evaluation on adaptive condition relaxation and creation of suspiciously empty regions, in our response to Reviewer [Ekyx]. With this cluster-wise correlation structure, we can emulate a dataset with "higher-order or context-dependent" interactions between variables.

The table below shows the comparison results of KASS (k=2k=2 and η=1\eta=1) against SMOTE and TabSyn on these small-sample, higher-order interaction datasets. The average values and standard deviations across 10 iterations are reported. Indeed, KASS exhibits a high restart rate (around 0.69) for the dataset with a sample size of N=100N=100. Additionally, KASS shows suboptimal data utility performance with respect to column similarity and the Alpha-Beta score compared to SMOTE and TabSyn. This reflects the difficulty of sampling privacy-preserving synthetic data under a limited sample size. Nevertheless, KASS still succeeds in preventing privacy leakage, yielding zero UiOiS counts. This demonstrates that the utility-privacy tradeoff still holds in the small-sample regime. The restart rate drops drastically (to around 0.13) as we increase the sample size to N=500N=500, which is still small compared to typical datasets used in synthetic data generation. Furthermore, the data utility metrics of KASS become comparable to those of TabSyn, although SMOTE remains optimal in terms of data utility. However, this comes at the cost of replicating rare values, as evidenced by its high UiOiS counts, an issue that KASS successfully avoids.

NModelRestart rateCol-SimAlpha-BetaAuth.UiOiS
100KASS (k=2,η=1k=2, \eta = 1)0.6950±0.10520.6950 \pm 0.10520.8988±0.01810.8988 \pm 0.01810.3884±0.03360.3884 \pm 0.03360.5810±0.06950.5810 \pm 0.06950.00±0.000.00 \pm 0.00
SMOTE-0.9214±0.01150.9214 \pm 0.01150.6487±0.07440.6487 \pm 0.07440.3960±0.05500.3960 \pm 0.055010.50±3.5410.50 \pm 3.54
TabSyn-0.9234±0.01320.9234 \pm 0.01320.4399±0.05780.4399 \pm 0.05780.5170±0.05400.5170 \pm 0.05400.80±0.920.80 \pm 0.92
500KASS (k=2,η=1k=2, \eta = 1)0.1320±0.01620.1320 \pm 0.01620.9552±0.00580.9552 \pm 0.00580.6550±0.01570.6550 \pm 0.01570.5036±0.02470.5036 \pm 0.02470.00±0.000.00 \pm 0.00
SMOTE-0.9640±0.00480.9640 \pm 0.00480.8194±0.01740.8194 \pm 0.01740.3862±0.02750.3862 \pm 0.027579.00±7.1579.00 \pm 7.15
TabSyn-0.9618±0.00780.9618 \pm 0.00780.6614±0.02050.6614 \pm 0.02050.5186±0.02510.5186 \pm 0.02513.40±1.433.40 \pm 1.43

To evaluate the robustness of the pairwise NMI thresholding strategy on datasets with higher-order interactions, we compute the cluster-wise joint TVD and removal rate, which measure the preservation of joint correlations and the proportion of correlated features removed, respectively. For a detailed description of each metric, please refer to the section, Evaluation on adaptive condition relaxation and creation of suspiciously empty regions, in our response to Reviewer [Ekyx]. Clearly, the removal rate is high (around 0.33), especially within the minority cluster when generating the N=100N=100 dataset. This highlights a limitation of KASS under complex dependence structures. Nevertheless, accounting for standard deviations, KASS achieves a joint TVD closely aligned with those of other baseline models, except within the minority cluster for the N=500N=500 dataset. The joint TVD results reveal that even though KASS may inadvertently drop conditioning variables involved in higher-order interactions, the algorithm still manages to preserve the overall joint distribution.

We will discuss these findings in Section 3.2 and report the detailed simulation results in the appendix.

NModelClusterSample countsJoint TVDRemoval rate
100KASS (k=2,η=1k=2,\eta = 1)172.90±4.0972.90 \pm 4.090.1638±0.05030.1638 \pm 0.05030.2102±0.01700.2102 \pm 0.0170
KASS (k=2,η=1k=2,\eta = 1)227.10±4.0927.10 \pm 4.090.1546±0.07510.1546 \pm 0.07510.3309±0.01090.3309 \pm 0.0109
SMOTE181.90±3.7881.90 \pm 3.780.1354±0.03670.1354 \pm 0.0367-
SMOTE218.10±3.7818.10 \pm 3.780.2020±0.12950.2020 \pm 0.1295-
TabSyn179.50±3.8979.50 \pm 3.890.1483±0.03560.1483 \pm 0.0356-
TabSyn220.50±3.8920.50 \pm 3.890.1335±0.05370.1335 \pm 0.0537-
500KASS (k=2,η=1k=2,\eta = 1)1391.10±8.91391.10 \pm 8.910.0927±0.02400.0927 \pm 0.02400.1492±0.00730.1492 \pm 0.0073
KASS (k=2,η=1k=2,\eta = 1)2108.90±8.91108.90 \pm 8.910.1916±0.03010.1916 \pm 0.03010.2486±0.00790.2486 \pm 0.0079
SMOTE1397.70±6.99397.70 \pm 6.990.0709±0.02340.0709 \pm 0.0234-
SMOTE2102.30±6.99102.30 \pm 6.990.0937±0.05320.0937 \pm 0.0532-
TabSyn1397.60±10.07397.60 \pm 10.070.0777±0.01620.0777 \pm 0.0162-
TabSyn2102.40±10.07102.40 \pm 10.070.0924±0.02950.0924 \pm 0.0295-

Adding MIA and AIA for privacy measure (Q3)

Please refer to responses Membership inference attacks, Attribute inference attacks, and Subgroup analysis of inference attacks to the reviewer [Q2vE] who raised the same concern.

Thank you for the response and additional experiments. However, my concerns about small-sample robustness and higher-order dependency preservation remain, particularly given the high removal rate in the minority cluster at N=100N=100 and its higher Joint TVD compared with SMOTE and TabSyn at N=500N=500. I therefore maintain my original score.

We appreciate your careful reassessment of our paper and your constructive feedback throughout the discussion. Your comments have been valuable in helping us improve the quality of the manuscript. Thank you again for your time and thoughtful consideration.

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

The paper proposes a new sequential sampling framework for generation of realistic tabular synthetic data, named KASS, which embeds k-anonymity and mutual information in the sampling process. The experiments cover several kinds of tabular datasets (some experiments only presented in the appendix) and establish that the proposed method is competitive with (confidence interval not provided) existing approaches, while also being faster than the considered deep learning systems. The manuscript is well written, contains most of the key references you would expect for a contribution on realistic synthetic data generation, and the mathematical formalisation is well worked and consistent throughout the main matter. I tend towards recommending acceptance, with some minor conditionals, namely that confidence intervals (error bars) are added to the major experiment results. I recommend that the high-dimensional dataset experiment (Appendix C5) is worked into the main matter, and that perhaps the differential privacy connection is dialed back or omitted, (e.g., save for future dedicated DP-KASS paper), and that the authors consider an empirical privacy evaluation also (e.g. simulated membership inference attack, or epsilon identifiability risk).

Critical strengths
  1. Sequential synthesis with classical models is often dismissed as a viable solution in the synthetic data community even though they often match or outperform deep learning solutions on utility. This work proposes a new sequential generation procedure that also place well on the benchmarks, potentially offering new directions and reigniting this important debate.
Critical weaknesses
  1. Claiming “competitive utility and markedly outperforms competing methods” without confidence intervals or rank analysis is a major flaw in the experimental design and analysis, especially when the settlement often comes down to something happening in the third or fourth decimal place (see tables in appendix). Apparently, all numbers are averages of 10 measurements, so standard error measures or standard deviation numbers should be provided.
  2. The introduction and related works have a weird pacing, you have to go to page 3 to before privacy preserving synthetic data generation with differential privacy is mentioned; up until that point it is all “other methods are bad for not having baked in privacy, we need this new thing”. Arguing the need for a new generation of smart phones based off that your dated TI-83 graphing calculator cannot make a phone call is really a false analogy fallacy.
  3. There is a lot of moving parts, fallbacks, and special rules worked into the proposed methodology. That all ends up looking like a roundabout way of achieving what sequential decision tree models are already doing. Similarly, I get that reviewers on these kinds of papers always ask about differential privacy (believe me we have been there ourselves), but I do not appreciate 11 pages of challenging proof content in the appendix that serve to establish a fuzzy connection with the concept. There is no such thing as approximately differentially private, either a method fully adheres to differential privacy (if so, it is a main feature of the contribution) otherwise it does not.
Questions for authors

k-anonymity has some strengths in its simplicity, but also give rise to some concerns. Notably that it has been proven vulnerable to a series of attacks You mention bias aversion in line 269, but if the relaxation drops conditioning when it would otherwise have become grid-locked, how do we avoid another bias in some samples caused by this fallback? It seems ill advised that some samples are generated in a different manner than others, what happens when you compare the “naturally conditioned” and the “relaxed” samples? In the example of Figure 1, imagine that the imbalance between the two categories of X3 were more pronounced, then picking the minority at an early stage of generation would (be likely to) always trigger a relaxation later on; effectively this would provoke a separation of the resulting generated data space, where majority class cases follow a different underlying probabilistic structure in the later variables than for the minority (independent) minority case. I understand that random sample-wise ordering could conceptually smear this effect out on average, but I think that there could still be some fail states for low-dimensional data and skewed variables. The sequential CART models (e.g., synthpop) use information gain to construct tree models based off variables available at that step in the synthesis chain, predicting the next variable; it uses the more relevant variables, and may decide to omit others. With hyperparameters we can control pruning, tree depth, and constrain the minimum number of samples of each outcome that will be in a terminal node. There seem to be many elements of the sequential tree models that are functionally equivalent to your approach, moreover they perform fairly similar in the comparison. One key distinction is that once a sequential tree model is trained; I can draw any number of new samples from it, knowing that all of them will follow the same (learned) distribution, however in you model two samples may be produced by vastly different random processes -
With the post hoc approaches that remove synthetic samples that are “too close” in the sample space, one risk that happens if data is greatly oversampled, is areas in the dataset space that are unnaturally vacuous, which lets an adversary identify that a sensitive real sample is “shadowing” that space. With (line 136) “Any value supported by fewer than k matching records, but at least one, is flagged as a privacy risk and removed from the sampling distribution” do that not run the risk of those same suspiciously vacant areas of the dataset distribution? I would recommend you get rid of the differential privacy connection (section 3.3 and appendix G) if it is not fully established (line 377; “KASS is not a DP synthesizer”). As I mention above, my experience with DP is that there is no such thing as halfway-DP, either you have it in full, or some assumption has been violated somewhere and your guarantee is void. If this is a route you want to go in the future save it for that and do “DP-KASS” as a dedicated project.
It would have been nice to see some empirical privacy measures also in your evaluation. MIA and epsilon identifiability, for instance, are fairly standard in these Non-DP private synthetic data generation papers, and it feels a bit shallow to only measure privacy from a perspective where we would expect KASS to excel (UiOiS) and not generally. You should spell out clearly somewhere (e.g. in the limitations) that the model needs access to the real data at training time, which makes a trained model unsuitable for publication; which is a weakness compared with end-to-end privacy frameworks such as a properly built differentially private model.

5 comments

Variability and performance comparisons

Figure 2 in the original manuscript reports averages over ten synthetic datasets, with standard deviations shown as error bars. Some of these bars are difficult to distinguish due to the small standard deviations. Updated results including GReaT are provided in the supplementary material (see the README file). We will report standard deviations in all appendix result tables and interpret small differences cautiously.

Revision of the privacy analysis

Our original aim was to use the DP framework to characterize the privacy properties of KASS beyond the selected empirical disclosure metrics. However, as discussed in our response to Reviewer Q2vE, the DP bound stated in Theorem 1 does not hold for the implemented sampling procedure. Following your recommendation, we will remove the DP analysis in Section 3.3 and Appendix G, along with the associated claims throughout the manuscript, to avoid any confusion.

Revisions: Introduction, Related Work, and comparison with Synthpop

We will revise the Introduction and Related Work to clarify the positioning of KASS. DP synthesizers provide formal privacy guarantees, but the associated utility loss can limit their practical use. Non-DP deep generative models offer flexible modeling of complex dependencies. However, selecting parameters such as model size to meet an empirical disclosure criterion can require repeated training and generation at substantial computational cost.

Tree-based sequential methods such as Synthpop can offer efficient synthesis with competitive utility, although conditional tree fitting can become expensive in high-dimensional settings. They may also reproduce original records. Pruning can reduce this risk, but reproduction restrictions under a fixed synthesis order may leave parts of the distribution unrepresented. Our fixed-order experiment below illustrates this concern under KASS's support restrictions.

KASS therefore combines low-support candidate restrictions with a randomized order for each generated record. Empirical conditional counts accommodate varying orders and conditioning subsets without fitting separate conditional models, avoiding this fitting overhead in high-dimensional settings. To limit utility loss during relaxation, only conditions with pairwise NMI below the selected threshold are eligible for removal. Although pairwise NMI does not explicitly assess higher-order dependencies, we expect such dependencies to be reflected in generation when the corresponding patterns have sufficient support and the relevant conditions are retained. This design reduces computational cost by avoiding exhaustive evaluation of higher-order dependencies.

We will also clarify that the fitted KASS generator cannot be publicly released without exposing the original training data. In addition, we will move the high-dimensional experiment in Appendix C5 into the main text.

Adding MIA and AIA for privacy measure

Please refer to responses Membership inference attacks, Attribute inference attacks, and Subgroup analysis of inference attacks to the reviewer [Q2vE] who raised the same concern.

Evaluation on adaptive condition relaxation and creation of suspiciously empty regions

To evaluate the concerns regarding the sampling bias for low-dimensional and highly imbalanced (or skewed) variables, we generate mixed-type simulation datasets of size N∈100,500N \in \\{100, 500\\} partitioned into two imbalanced clusters $c \in \{1, 2\}$ at an 8:28:2 ratio, (n1,n2)∈(80,20),(400,100)(n_1, n_2) \in \\{(80, 20), (400, 100)\\}, with eight features X=(X1,…,X8)\mathbf{X} = (X_1, \dots, X_8) via a Gaussian copula. That is, the datasets consist of total 9 variables including the cluster information. Each cluster has an active variable subset AcA_c (A1=1,2,3,4,5A_1 = \\{1, 2, 3, 4, 5\\}, A2=1,2,6,7,8A_2 = \\{1, 2, 6, 7, 8\\}), where X1,X2\\{X_1, X_2\\} are shared and the remaining active variables are cluster-specific. We draw Z∼N(0,Σ(c))\mathbf{Z} \sim \mathcal{N}(\mathbf{0}, \boldsymbol{\Sigma}^{(c)}) with correlation Σj,k(c)=ρ=0.85\Sigma^{(c)}_{j,k} = \rho = 0.85 if j≠k∈Acj \neq k \in A_c (and 00 otherwise), transform it into uniform latent variables Uj=Φ(Zj)U_j = \Phi(Z_j), and apply inverse CDFs to obtain three variable types:

X_1, X_2 &\sim \text{Categorical}(\\{\text{A}, \text{B}, \text{C}\\}, \mathbf{p}^{(c)}), && \mathbf{p}^{(1)} = (0.90, 0.08, 0.02), \quad \mathbf{p}^{(2)} = (0.02, 0.08, 0.90), \\\\ X_3, X_4, X_6, X_7 &\sim \text{LogNormal}(\mu_{c,j}, \sigma^2), && \sigma = 0.35, \quad \mu_{c,j} = 3.5 \cdot \mathbb{I}\\!\left(j \in A_c\right), \\\\ X_5, X_8 &\sim \text{Geometric}(p_{\text{geom}}), && p_{\text{geom}} = 0.05 \text{ on } \\{1, 2, \dots\\}. \end{aligned}$$ Here, the cluster variable can be understood as critical information that impacts the distribution and correlation structure of other variables, such as nationality in demographic data. Now, using a fixed sampling order of $[\text{Cluster}, X_4, X_8, X_2, X_1, X_3, X_5, X_7, X_6]$ and parameter settings of $k=2$ and $\eta = 1$, we generated 10 synthetic datasets for each dataset size $N \in \\{100, 500\\}$. When the minority cluster ($c=2$) is sampled first, we can expect difficulties in sampling the subsequent variables due to the $k$-anonymity constraint. Consequently, the algorithm more frequently relaxes the conditional variables or restarts compared to when the majority cluster is sampled. In addition, synthetic datasets with a random sampling order were also generated to evaluate the effect of randomizing the sampling order. The table below shows the cluster-wise summary of key statistics for the generated synthetic datasets, with average values and standard deviations reported. | $N$ | Sampling order | Cluster | Sample counts | Joint TVD | Removal rate | |---|---|---|---|---|---| | 500 | Fixed | 1 | $439.50 \pm 7.18$ | $0.2726 \pm 0.0266$ | $0.2978 \pm 0.0033$ | | | Fixed | 2 | $60.50 \pm 7.18$ | $0.4785 \pm 0.0336$ | $0.0147 \pm 0.0110$ | | | Random | 1 | $391.10 \pm 8.91$ | $0.0927 \pm 0.0240$ | $0.1492 \pm 0.0073$ | | | Random | 2 | $108.90 \pm 8.91$ | $0.1916 \pm 0.0301$ | $0.2486 \pm 0.0079$ | | 100 | Fixed | 1 | $90.00 \pm 2.11$ | $0.3983 \pm 0.0400$ | $0.3008 \pm 0.0092$ | | | Fixed | 2 | $10.00 \pm 2.11$ | $0.9500 \pm 0.0000$ | $0.2667 \pm 0.0000$ | | | Random | 1 | $72.90 \pm 4.09$ | $0.1638 \pm 0.0503$ | $0.2102 \pm 0.0170$ | | | Random | 2 | $27.10 \pm 4.09$ | $0.1546 \pm 0.0751$ | $0.3309 \pm 0.0109$ | In the table, in addition to the generated sample counts, we present two evaluation metrics, joint TVD (total variation distance) and removal rate. For each cluster $c \in \\{1, 2\\}$, let $\tilde{A}_c = \\{X_j : j \in A_c\\} \cup \\{c\\}$ be its $m = 6$ total active variables. To compute joint TVD, numeric features are binarized at their original full sample medians ($\tilde{X}_j = \mathbb{I}(X_j > m_j)$), and we measure the total variation distance between the empirical joint distributions $\hat{P}_c$ (original) and $\hat{Q}_c$ (synthetic) over $\tilde{A}_c$: $\text{Joint TVD}_c = \frac{1}{2} \sum_{\mathbf{v}} |\hat{P}_c(\mathbf{v}) - \hat{Q}_c(\mathbf{v})|$. This metric evaluates how faithfully the synthetic data preserves the true within-cluster joint dependency structure among active variables (lower values indicate higher fidelity). To compute the removal rate, for each synthetic record $i \in S_c$ with active variables ordered by generation as $(V_{i,1}, \dots, V_{i,m})$, let $P_{i,r} = \\{V_{i,1}, \dots, V_{i,r-1}\\}$ be the preceding active variables at step $r$ and $K_{i,r}$ be the retained conditions; the fraction of dropped active conditions across $\binom{m}{2}=15$ opportunities per record is $\text{Removal rate}_c = \frac{1}{|S_c|\binom{m}{2}} \sum_{i \in S_c} \sum_{r=1}^{m} |P_{i,r} \setminus K_{i,r}|$. This metric quantifies how frequently active conditioning variables are relaxed during sequential sampling to satisfy the $k$-anonymity threshold within each cluster.

As expected, with a fixed sampling order, KASS generates significantly fewer samples for the minority cluster than indicated by the original ratio. This demonstrates that if KASS begins by sampling a minority class in a highly imbalanced variable, the algorithm restarts with high probability due to the privacy constraint. Furthermore, even when KASS succeeds in sampling minority cluster records, the resulting synthetic records exhibit severe sampling bias, as evidenced by the joint TVD values. Under the fixed sampling order, the joint TVD within the minority cluster is higher than that within the majority cluster for both the N=100N=100 and N=500N=500 cases. Surprisingly, removal rates are higher for the majority class under fixed-order sampling. This occurs because the removal rate can only be measured for samples that did not undergo the restart procedure, and such samples follow a specific generation path that avoids restarting (see the FAMD visualizations in the supplementary material). Hence, with a fixed sampling order, KASS can indeed suffer from sampling bias caused by imbalanced or skewed variables.

However, with a random sampling order, we obtain substantially better results for both the majority and minority clusters. Clearly, KASS generates minority cluster samples in appropriate proportions, and the joint TVD drops considerably for both clusters. For the dataset of size N=100N=100, the joint TVD values are comparable between Clusters 1 and 2. This experiment confirms that randomizing the sampling order is critical for mitigating the sampling bias stemming from imbalanced or skewed variables.

The effectiveness of the random sampling order design also holds when investigating suspiciously empty regions in the generated sample space. Figures 1 and 2 present visualizations of the original and selected synthetic datasets using FAMD, a dimensionality reduction technique for mixed-type data (see the corresponding figures in the supplementary material). Note that orange and blue samples represent the minority and majority clusters, respectively. For both sample sizes, synthetic samples belonging to the minority class are concentrated in a specific region when generated with fixed-order sampling (ten synthetic samples were generated for the N=100N=100 dataset, but only two points are visible due to overlap). This leaves "suspiciously empty" regions in the sample space that were occupied by samples in the original dataset. An attacker with prior knowledge of the distributional characteristics of the minority class could exploit this spatial distortion to infer information about rare and private samples in the original dataset. However, when a random sampling order is used, this phenomenon substantially diminishes, as shown in the right subplots. The original sample space of the minority cluster is largely recovered, as is that of the majority class. These visualization results further underscore the importance of a random sampling order in overcoming the limitations KASS may face when dealing with datasets containing rare samples.

We will expand the discussion of random sampling order in Section 3.2 and report the detailed simulation results and visualizations in the appendix.

Reviewer_Q2vEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes KASS, a sequential sampler for synthetic tabular data. It excludes candidate values with positive empirical counts below k, relaxes sparse conditioning sets by dropping weakly associated variables according to normalized mutual information, and restarts when relaxing a strongly dependent condition would be necessary. A restricted-set depth controls how many earlier exclusions are retained. Continuous variables are discretized and reconstructed after sampling. Experiments on five main datasets, together with supplementary ablations and a sparse high-dimensional dataset, examine downstream utility, constraint violations, distributional quality, unique-record reproduction, and runtime. The paper also analyzes simplified and conditional privacy bounds.

Critical strengths
  1. The sampling procedure is direct and its components have understandable roles. Low-count filtering targets rare-record reproduction, dependence-aware relaxation addresses sparse conditional supports, and the restart rule limits the removal of strongly associated variables. The design makes the utility, feasibility, and disclosure-risk trade-offs easier to inspect than an opaque end-to-end generator (Sections 3.1-3.2).
  2. The experiments support several specific design choices. On Student, a fixed sampling order reduces weighted F1 from 0.7483 to 0.4025, while removing the restart rule increases average constraint violations from 0 to 42.5 with little change in predictive utility (Appendix C.3, Table 2). This usefully shows why downstream accuracy alone is insufficient. The main datasets take approximately 3.6-40 seconds to generate, and the Insurance experiment provides additional evidence of practical utility and efficiency on sparse data.
Critical weaknesses
  1. Language-model-based tabular synthesizers are missing. The baselines include non-deep methods (Synthpop, SMOTE, AIM, and MST), but omit GReaT [1], TabuLa [2], and DP-LLMTGen [3]. A representative comparison or substantive exclusion rationale is needed to support broad competitiveness. Utility, constraint validity, disclosure risk, and total cost should use consistent preprocessing, while distinguishing privacy assumptions: a k-support restriction and DP cannot be equated through similarly named parameters.
  2. Privacy metrics do not cover the most relevant attacks. Zero UiOiS establishes non-reproduction of original unique records under the chosen discretization; neither it nor distance-based authenticity rules out membership or sensitive-attribute inference. Adaptive conditioning may leak information without copying complete rows. Attack evaluations need held-out controls, rare-subgroup analysis, and discretization sensitivity. Current evidence supports reduced exact reproduction, rather than general anonymity (Section 4; Appendix F).
  3. Adaptive conditioning limits the anonymity interpretation. Dropped variables remain in completed records, whose full combinations need not have k supporting records, as Section 3.3 acknowledges. The adaptive analysis also assumes unenforced boundary-support diversity and excludes parts of the discretization, NMI, and reconstruction pipeline. These limitations do not negate practical utility, but require framing the guarantee as a sampling-time restriction rather than full-record k-anonymity or end-to-end DP.
  4. Theorem 1 appears inconsistent with actual sampling. After filtering, remaining counts must be normalized by their sum; Appendix G.3 (p. 30) instead uses the thresholded parent count. For one variable with k=5 and 100 records (five a's, five b's, 90 distinct singletons), filtering gives P(a)=1/2. Removing one a excludes a, giving P(a)=0. Under the appendix's add/remove adjacency, approximate DP therefore requires delta at least 1/2, regardless of epsilon, contradicting delta=k/N, approximately 0.05. This requires no adaptive conditioning. A mechanism retaining an explicit failure outcome without renormalization would need separate specification and reconciliation with the sampler returning complete records.
Questions for authors
  1. Could you provide the correctly normalized single-record distribution and explain how Theorem 1 handles the one-variable example in Weakness 4? Please distinguish any explicit failure mechanism from rejection or restart until a valid record is returned.
  2. What exact protection is guaranteed by the adaptive sampler for a completed record, and is the boundary-support condition checked or enforced? How should users interpret k once conditioning variables are dropped?
  3. Could you add membership and attribute-inference evaluations with held-out controls, together with a sensitivity analysis for discretization and rare subgroups?
  4. Could you compare with a representative language-model-based synthesizer, such as GReaT or DP-LLMTGen, reporting utility, constraints, attack-based disclosure metrics, and total cost while explicitly separating their privacy assumptions?
7 comments

Revision of the privacy analysis (Q1, Q2)

Our original aim was to use the DP framework to characterize the privacy properties of KASS beyond the selected empirical disclosure metrics. We acknowledge that the reviewer's counterexample is valid and that the derivation of Theorem 1 does not accurately reflect the implemented sampling procedure. The implemented sampler draws from retained candidates with probabilities proportional to their counts and retries unsuccessful generation attempts, returning an error after repeated failures. Moreover, additional conditions, including the boundary-support condition, were imposed for the theoretical analysis but are not enforced in the implementation. Although we believe the theorem and its proof could be revised to address the identified issues, we have decided to follow Reviewer Ekyx's suggestion to remove the DP analysis unless a stronger theoretical guarantee can be established. Accordingly, we will remove Section 3.3 and Appendix G, along with the associated claims throughout the manuscript, to avoid confusion about the privacy guarantees of KASS.

We will also clarify the roles of kk and η\eta in KASS. In its internal discrete representation, KASS relaxes conditioning by using smaller subsets of previously generated variables when necessary, while preserving previously identified exclusions of low-support values. During this relaxation process, larger values of η\eta increase the tendency to suppress value combinations that are sparsely represented in the original training data. Consequently, each successfully generated record in this representation is either absent from the original training data or appears at least kk times. However, these restrictions do not establish differential privacy or kk-anonymity for the released synthetic data. In particular, kk-anonymity requires each released record to share its quasi-identifier values with at least k−1k-1 other released records.

Finally, we will incorporate the membership inference attack (MIA) and attribute inference attack (AIA) evaluations presented below into the revised manuscript and discuss the privacy limitations of KASS in light of these results.

Comparison with GReaT (Q4)

We additionally evaluate GReaT with the DistilGPT-2 backbone used in TabSyn's baseline implementation. We use 100 training epochs, a batch size of 32 (16 on Student due to GPU memory limits), and a maximum generation length of 512 tokens. As with the other baselines, we train the model once per dataset and generate ten synthetic datasets. We discard records containing categorical values absent from the training data and continue sampling until the synthetic dataset matches the size of the original training set. Time includes training and the average generation time for one synthetic dataset.

DatasetModelCol-Sim ↑Pair-Sim ↑Weighted-F1 ↑Violations ↓Alpha–Beta ↑Auth. ↑UiOiS ↓Time (s) ↓
AdultKASS0.9896±0.00060.9896 \pm 0.00060.9811±0.00270.9811 \pm 0.00270.8634±0.00250.8634 \pm 0.002500.7487±0.00140.7487 \pm 0.00140.4930±0.00210.4930 \pm 0.0021047.247.2
GReaT0.9246±0.00050.9246 \pm 0.00050.8182±0.02700.8182 \pm 0.02700.8588±0.00230.8588 \pm 0.00232.6±1.72.6 \pm 1.70.7145±0.00100.7145 \pm 0.00100.5131±0.00280.5131 \pm 0.00281742.7±35.91742.7 \pm 35.917 045.417\,045.4
DefaultKASS0.9581±0.00070.9581 \pm 0.00070.9068±0.02640.9068 \pm 0.02640.7834±0.00450.7834 \pm 0.0045–0.7486±0.00240.7486 \pm 0.00240.4889±0.00330.4889 \pm 0.0033025.725.7
GReaT0.8261±0.00170.8261 \pm 0.00170.7322±0.01850.7322 \pm 0.01850.7965±0.00370.7965 \pm 0.0037–0.6567±0.00220.6567 \pm 0.00220.5263±0.00360.5263 \pm 0.0036212.2±11.4212.2 \pm 11.415 807.115\,807.1
FoodKASS0.9726±0.00140.9726 \pm 0.00140.9399±0.00740.9399 \pm 0.00740.5740±0.00760.5740 \pm 0.0076478.7±20.8478.7 \pm 20.80.7748±0.00210.7748 \pm 0.00210.4771±0.00350.4771 \pm 0.003503.93.9
GReaT0.9129±0.00160.9129 \pm 0.00160.9345±0.01240.9345 \pm 0.01240.5171±0.00760.5171 \pm 0.0076282.4±11.9282.4 \pm 11.90.7237±0.00850.7237 \pm 0.00850.5203±0.01060.5203 \pm 0.01061416.8±29.31416.8 \pm 29.34 569.24\,569.2
ShoppersKASS0.9317±0.00110.9317 \pm 0.00110.9575±0.00210.9575 \pm 0.00210.8883±0.00690.8883 \pm 0.006911.4±2.611.4 \pm 2.60.7080±0.00290.7080 \pm 0.00290.4544±0.00480.4544 \pm 0.004806.16.1
GReaT0.8572±0.00060.8572 \pm 0.00060.8782±0.00800.8782 \pm 0.00800.8922±0.00600.8922 \pm 0.00600.8±0.80.8 \pm 0.80.5961±0.02560.5961 \pm 0.02560.5285±0.00600.5285 \pm 0.0060165.4±9.5165.4 \pm 9.55 416.65\,416.6
StudentKASS0.9791±0.00130.9791 \pm 0.00130.9427±0.00930.9427 \pm 0.00930.7483±0.01710.7483 \pm 0.017100.7887±0.00400.7887 \pm 0.00400.4167±0.00850.4167 \pm 0.008504.74.7
GReaT0.9317±0.00170.9317 \pm 0.00170.8217±0.01910.8217 \pm 0.01910.7406±0.00570.7406 \pm 0.005719.0±4.219.0 \pm 4.20.7046±0.00570.7046 \pm 0.00570.4948±0.01150.4948 \pm 0.011505 292.85\,292.8

The table shows the results for GReaT with KASS (k=2,η=1k=2,\eta=1) across the five datasets. Compared with GReaT, KASS achieves higher average scores for Column Similarity (Col-Sim), Pairwise Similarity (Pair-Sim), and Alpha-Beta on all five datasets. For weighted-F1, the two methods achieve similar average scores on most datasets, but KASS performs better on Food. GReaT's higher authenticity scores may partly reflect its lower distributional fidelity. KASS also yields zero UiOiS counts on every dataset, whereas GReaT yields nonzero counts on four datasets. Additional comparisons with the other deep learning baselines show that GReaT yields the highest UiOiS counts on Adult, Default, and Shoppers (see the README file in the supplementary material). GReaT produces the fewest constraint violations among the deep learning baselines on all four datasets with predefined constraints. This may reflect the language model's ability to capture constraint-related dependencies between attributes. However, GReaT requires substantially more time than all other methods across the five datasets. We will incorporate these findings into the main experimental section of the revised manuscript.

Membership inference attacks (Q3)

We adapt the Classifier baseline evaluated in Ensemble-MIA[1], using a scikit-learn random forest. Numerical standardization and categorical one-hot encoding are fitted on each synthetic dataset. We train the attacker to distinguish synthetic records from a fixed reference set consisting of half of the held-out test data. The same split is used across methods and generation draws. The attack is evaluated on original training records (members) and the remaining test records (non-members). We use the predicted probability of being synthetic as the attack score and measure attack performance by AUROC.

We compare four KASS configurations (k=2,5k=2,5 and η=1,3\eta=1,3) with TabSyn, GReaT, SMOTE, and Synthpop on the five benchmark datasets. For each configuration and dataset, we generate ten synthetic datasets from a single fitted model and report the average and standard deviation of AUROC.

ModelAdultDefaultFoodShoppersStudent
KASS (k=2,η=1k=2,\eta=1)0.5610±0.00160.5610 \pm 0.00160.5739±0.00270.5739 \pm 0.00270.5516±0.00250.5516 \pm 0.00250.5819±0.00910.5819 \pm 0.00910.6043±0.00700.6043 \pm 0.0070
KASS (k=2,η=3k=2,\eta=3)0.5408±0.00380.5408 \pm 0.00380.5327±0.00210.5327 \pm 0.00210.5539±0.00400.5539 \pm 0.00400.5544±0.00310.5544 \pm 0.00310.5763±0.00760.5763 \pm 0.0076
KASS (k=5,η=1k=5,\eta=1)0.5471±0.00300.5471 \pm 0.00300.5766±0.00310.5766 \pm 0.00310.5168±0.00230.5168 \pm 0.00230.5808±0.00520.5808 \pm 0.00520.6150±0.01690.6150 \pm 0.0169
KASS (k=5,η=3k=5,\eta=3)0.5150±0.00360.5150 \pm 0.00360.5275±0.00380.5275 \pm 0.00380.5162±0.00360.5162 \pm 0.00360.5423±0.00420.5423 \pm 0.00420.5734±0.01410.5734 \pm 0.0141
TabSyn0.5035±0.00240.5035 \pm 0.00240.4945±0.00410.4945 \pm 0.00410.5187±0.00410.5187 \pm 0.00410.5052±0.00820.5052 \pm 0.00820.5032±0.01940.5032 \pm 0.0194
GReaT0.5008±0.00210.5008 \pm 0.00210.4936±0.00270.4936 \pm 0.00270.5053±0.00150.5053 \pm 0.00150.5221±0.00300.5221 \pm 0.00300.5140±0.00810.5140 \pm 0.0081
SMOTE0.6161±0.00310.6161 \pm 0.00310.5880±0.00450.5880 \pm 0.00450.5765±0.00340.5765 \pm 0.00340.6342±0.00410.6342 \pm 0.00410.6609±0.01170.6609 \pm 0.0117
Synthpop0.5499±0.00360.5499 \pm 0.00360.5152±0.00400.5152 \pm 0.00400.5485±0.00630.5485 \pm 0.00630.5493±0.00830.5493 \pm 0.0083–

KASS generally yields higher average MIA AUROC than TabSyn and GReaT, but lower AUROC than SMOTE. At k=2,η=1k=2,\eta=1, its average AUROC is also higher than Synthpop's on all four datasets with available results. Although KASS achieves zero UiOiS in the main experiments, these attack results indicate residual susceptibility to membership inference. Increasing η\eta from 1 to 3 generally lowers average AUROC, with little change on Food, suggesting that parameter adjustment can help reduce this susceptibility. We also note that SMOTE achieves the highest average AUROC among the reported methods on all five datasets. This can be interpreted as reflecting its interpolation between neighboring training records, which may make members easier to distinguish from non-members.

[1] Ensembling Membership Inference Attacks Against Tabular Generative Models, AISec 2025.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.