Synthetic Data Generator (SDG)

Updated 23 November 2025

Synthetic Data Generator (SDG) is a method that creates artificial datasets replicating the statistical, structural, and causal properties of real data.
It employs diverse paradigms including marginals-based models, causal generative models, and rule-adhering mechanisms to ensure fidelity and fairness.
SDGs enable privacy-preserving analysis, rigorous system benchmarking, and fairness evaluation while balancing data utility with differential privacy constraints.

Synthetic Data Generator (SDG) refers to any algorithmic system that produces artificial datasets whose statistical, structural, or causal properties are modeled after real data, with the goal of mimicking the distributional, privacy, or domain-specific constraints of the original dataset. SDGs serve a broad array of tasks including privacy-preserving data analysis, system benchmarking, fairness investigations, and the augmentation of scarce or restricted datasets. SDG design and evaluation crosscut statistical modeling, privacy engineering, interpretability, and domain expertise, and the literature encompasses methods based on graphical models, causal inference, deep learning, privacy mechanisms, and more.

1. Fundamental Paradigms and Modeling Principles

The dominant SDG paradigms are rooted in distinct families of probabilistic modeling, causal generative modeling, rule-based data synthesis, and privacy-centric mechanisms. A unified view is that an SDG defines a (possibly randomized) mapping from either real data samples or pre-aggregated statistics to draws from a synthetic data distribution. Let $D$ represent the confidential dataset, $D'$ the synthetic output, and $\mathcal{M}$ the SDG mechanism. The structure of $\mathcal{M}$ falls into several classes:

Marginals-based SDGs: These select a collection of low-order marginals or conditional probabilities (e.g., 2- or 3-way tables), add noise as required (typically for privacy), and sample synthetic records so their statistics match the (privatized) measurements. Examples include MST, PrivBayes, and Private-GSD, which utilize maximum spanning trees, Bayesian networks, or genetic optimization to fit marginals (Golob et al., 7 Oct 2024).
Causal Generative Models (CGMs): CGMs employ a domain-informed causal graph, specifying structural equations for each variable as functions of parents and exogenous noise variables. These models, commonly formulated as Structural Causal Models (SCMs), enable the explicit modeling of interventions and fairness-related manipulations. For instance, in recruitment scenarios, separately structured CGMs for job-offers and candidate curricula, with domain-elicited edges and parametrizations, facilitate fairness-oriented synthetic data generation and downstream intervention experiments (Iommi et al., 20 Nov 2025).
Rule-Adhering SDGs: These incorporate domain rules (hard or soft constraints) either by post-filtering or by modifying the generative process directly. Auto-regressive neural samplers or GANs are regularized or structurally restricted to avoid or penalize generation of records violating expert-defined rules (Platzer et al., 2022).
Privacy-driven SDGs: Differentially private mechanisms (e.g., Laplace or Gaussian noise addition to marginals or regression statistics) guarantee formal privacy bounds $\varepsilon$ , $\delta$ . Marginal-based private SDGs, DP linear regression-generated SDGs, and multi-party computation (MPC)-backed collaborative SDGs are notable instantiations (Lin et al., 19 Oct 2025, Pentyala et al., 13 Feb 2024).
Hybrid/Custom Models: Efforts such as multi-part macro-to-microdata (e.g., GenSyn) combine dependency graphs, copula-based mixing, and maximum entropy corrections to synthesize microdata from aggregate statistics (Acharya et al., 2022).

2. Algorithmic Workflows and Sampling Procedures

SDGs follow diverse but composable algorithmic workflows, typically parameterized by the available data type (raw records, macro statistics), task constraints (fairness, rules), and privacy requirements. Common algorithmic templates include:

Marginals/PGM-based Synthetic Data Generation

Select–Measure–Generate:
- Selection: Identify query set $W$ (e.g., marginals, conditionals).
- Measurement: Query $W$ on $D$ , add calibrated noise per privacy mechanism.
- Generation: Fit a graphical model, generative neural net, or explicit distribution to match the measurements; sample new records (Pentyala et al., 13 Feb 2024, Golob et al., 7 Oct 2024, Maddock et al., 15 Apr 2025).
Causal SCM Generation:
- Structure Elicitation: Construct causal DAG via domain expert interviews.
- Mechanism Learning: Estimate conditional probabilities or functions for each node given parents.
- Sampling (Intervenable): Topologically order nodes; for each, sample exogenous noise and generate variable values per the structural equations. For fairness control, inject interventions (e.g., tilting $P(W_C|G)$ by an $\alpha$ parameter) (Iommi et al., 20 Nov 2025).
Rule-Adhering Generation:
- Constraint Injection: During training and/or sampling, enforce compliance with $k$ domain constraints via loss penalties and conditional probability masking.
- Sampling: Use autoregressive or GAN generator, reject or prohibit rule-violating draws (Platzer et al., 2022).

Differentially Private SDG Construction

Binning–Aggregation with Debiasing (for regression): Partition data space, privatize count/aggregate vectors by Gaussian noise, use bias-corrected estimators for sufficient statistics, and generate synthetic examples to mirror the DP estimator’s distribution (Lin et al., 19 Oct 2025).
Select–Measure–Generate-in-MPC: In distributed/multi-party settings, partitioned data holders secret-share local statistics, secure-MPC protocols aggregate and privatize, and the generator post-processes to synthesize new draws (Pentyala et al., 13 Feb 2024).

Macro-to-Micro SDGs

Graph & Copula Integration: Infer a dependency graph from available joint frequency tables, estimate conditionals, blend with a Gaussian copula capturing cross-location dependencies, and adjust synthetic profile weights via maximum entropy projection to fit univariate (and possibly higher-order) constraints (Acharya et al., 2022).

3. Fairness, Bias Control, and Interpretability

Modern SDGs are leveraged not only for data augmentation but also for controlled fairness evaluation and stress-testing.

Causal Fairness via Interventions: In structured CGMs, parameters (e.g., bias strength $\alpha$ for $G \to W_C$ ) can be systematically varied to simulate degrees of bias in sensitive pathways. Generated datasets enable measurement of fairness metrics such as Demographic Parity (DP) and Normalized Discounted Difference (rND) in downstream ranking tasks, with experiment design tracing how causal perturbations propagate through matching and ranking functions (Iommi et al., 20 Nov 2025).
Statistical Parity Fairness in Sampling: SDG label-generation can be post-processed by quantile-matching between privileged and unprivileged group score distributions, creating synthetic data yielding strong statistical parity across all thresholds. A tunable mixing parameter $\lambda$ supports utility–fairness tradeoffs, operationalized without retraining (Krchova et al., 2023).
Auditable SDGs: Some frameworks enforce decomposability, ensuring that SDG outputs depend only on pre-specified “safe” statistics and enabling ex post empirical audits to test for information leakage about protected patterns (Houssiau et al., 2022).

4. Privacy Properties, Attacks, and Limitations

SDG privacy is often defined via $(\varepsilon, \delta)$ -differential privacy (“distributional indistinguishability” for neighboring datasets), but vulnerabilities remain even in formal DP mechanisms:

Membership Inference (MI) Attacks: Algorithm-aware attacks (e.g., MAMA-MIA) exploit knowledge of SDG structure and selection to recover individual memberships or specific population statistics far more efficiently than black-box attacks, even in the presence of strong DP noise added to marginals (Golob et al., 7 Oct 2024).
Attribute Inference and Reconstruction: Linear reconstruction attacks, constructed as convex programs utilizing observed synthetic statistics, can accurately recover sensitive attributes for uniquely identified individuals unless privacy noise or model suppression sufficiently obscures these relationships. Larger synthetic data releases amplify inference risk despite improved utility (Annamalai et al., 2023).
Privacy–Utility Tradeoff: Increasing DP noise ( $\varepsilon \downarrow$ ) inhibits inference accuracy but reduces downstream data utility. Empirically, there exist regimes where neither high privacy nor high utility is achievable in non-DP or weak-DP SDGs (Pentyala et al., 13 Feb 2024, Annamalai et al., 2023).

5. Evaluation, Metrics, Applications, and System Design

Evaluations of SDGs are necessarily multi-faceted:

Statistical Fidelity: Compare synthetic versus real data via marginal, joint, or higher-order statistics: $L_1$ /KL/Wasserstein distances, mutual information matrices, and error on targeted query workloads (Golob et al., 7 Oct 2024, Acharya et al., 2022, Platzer et al., 2022).
Utility: Downstream task performance—held-out regression/classification $R^2$ , AUC, RMSE, cross-validation accuracy using real labels, as in TSTR (Train Synthetic, Test Real) and reciprocal schemes (Rousseau et al., 21 May 2025, Lin et al., 19 Oct 2025).
Privacy Metrics: Empirical privacy breach probabilities (e.g., MI attack AUC, TCAP), auditability via randomized generator cards, or DP budget simulations (Golob et al., 7 Oct 2024, Houssiau et al., 2022).
Fairness Metrics: Group fairness indices such as Demographic Parity (DP), rND, SPD curves, and the effect of intervention parameters (Iommi et al., 20 Nov 2025, Krchova et al., 2023).

SDG frameworks (e.g., SynthGuard (Brito et al., 14 Jul 2025)) now offer modular, auditable containerized platforms for defining, orchestrating, and enforcing privacy and compliance policies, integrating container workflow standards (Kubeflow/Argo/Kubernetes), multi-party cryptographic protocols, and comprehensive transparency via immutable audit logs.

Practical guidance includes: choosing SDG modeling complexity in accordance with data dimension/density, setting privacy parameters to balance statistical fidelity and reidentification risk, prioritizing explicit dependency of outputs on “safe” statistics for sensitive releases, and being alert to the specific avenues of privacy leakage exposed by each class of generator.

6. Advanced Topics, Frontiers, and Future Directions

Emerging research is pushing SDG boundaries in several directions:

Domain Specialization: Bespoke CGMs and hybrid workflows for recruitment, healthcare, or electric-vehicle energy modeling incorporate domain knowledge, causal pathways, expert rule sets, and rich memory-informed temporal structures (Iommi et al., 20 Nov 2025, Lahariya et al., 2022, Williams et al., 3 Sep 2024).
Macro-to-Micro Synthesis: Enabling microdata generation solely from aggregate public sources through dependency graph estimation, copula modeling, and maximum entropy projection (Acharya et al., 2022).
Text and Multimodal SDGs: LLM-driven frameworks encode time series and heterogeneous modalities as text for language-based synthesis, supporting highly efficient, small-sample, and multimodal conditional synthetic data outputs (Rousseau et al., 21 May 2025).
Vertical and Conditional SDG under DP: Exploit public–private attribute splits to optimize DP tradeoffs, using public attributes for exact model grounding and statistical noise on conditionals, or via conditional graphical model sampling (Maddock et al., 15 Apr 2025).
Collaborative SDG: Secure multi-party computation combined with DP enables distributed data holders to jointly generate synthetic data without trusting any single party with the complete real dataset (Pentyala et al., 13 Feb 2024).
Auditability and Decomposability: Frameworks formalize and empirically validate that generators depend only on user-specified statistical projections, supporting regulatory requirements for statistical “generator cards” and safe synthetic releases (Houssiau et al., 2022).

Limitations include scaling high-complexity models in high dimensions, bottlenecks in conditional sampling (when public attribute sets are large), quantifying privacy risk in novel modalities, and evolving regulatory standards.

7. Conclusion

Synthetic Data Generators are foundational to contemporary privacy-preserving data science, fairness benchmarking, and data-driven innovation under access restrictions. The field is characterized by a dynamic interplay of rigorous statistical modeling, domain-specific knowledge, privacy engineering, and empirical evaluation. Recent research delineates the limitations of purely marginal-based and standard DP mechanisms in guaranteeing privacy, emphasizes causally informed and interpretable generation, explores compositional privacy via multi-party and public–private attribute dimensions, and introduces formal auditing capabilities. The SDG research landscape continues to evolve, driven by expanding domains of application, adversarial privacy analysis, and increasing demands for transparency, fidelity, and regulatory compliance (Iommi et al., 20 Nov 2025, Lin et al., 19 Oct 2025, Maddock et al., 15 Apr 2025, Brito et al., 14 Jul 2025, Golob et al., 7 Oct 2024, Houssiau et al., 2022).