privacy-preserving data handling

This paper addresses three core challenges in privacy-preserving synthetic data generation (PP-SDG): opaque privacy loss interpretation, non-transparent risk semantics of the differential privacy parameter ε, and ambiguous, incomparable definitions across diverse privacy metrics (PMs). To resolve these, we systematically survey and, for the first time, formally unify the mathematical definitions of 17 mainstream PMs—explicitly specifying their underlying assumptions, implicit premises, and analytical expressions. Grounded in differential privacy theory, we integrate information-theoretic and statistical inference principles to analyze each PM’s computational model and applicability boundaries. Based on this analysis, we propose the first comprehensive PM taxonomy, rigorously characterized along three dimensions: completeness, consistency, and interpretability. This taxonomy substantially enhances transparency and standardization in privacy risk assessment and provides both a rigorous theoretical foundation and a practical evaluation framework for privacy–utility trade-offs in PP-SDG mechanisms.

This study addresses the dual challenges of privacy leakage from real user interaction data in AI training and insufficient representativeness of public datasets. We propose an end-to-end differentially private synthetic data generation framework for multimodal data (images, tabular, and text). Methodologically, it integrates sensitive data preprocessing, adaptive privacy budget allocation, generative-model-driven synthesis, and empirical privacy auditing, supporting both centralized and decentralized deployment. Our key contributions are: (1) the first systematic formulation of a synthetic data generation paradigm that jointly ensures distributional fidelity and rigorous (ε,δ)-differential privacy; (2) substantially improved cross-modal applicability and deployment trustworthiness; and (3) empirical validation demonstrating high data utility—even under stringent privacy constraints (ε ≤ 2)—enabling secure data sharing and reuse as a robust alternative to conventional anonymization techniques.

Data anonymization remains underutilized in practice due to its strong context dependency—requiring customization across data domains, protection units, scope, and privacy criteria—leading to a lack of general-purpose solutions; moreover, privacy guarantees of mainstream techniques (e.g., differential privacy) vary significantly across applications and are not universally equivalent. Method: Through empirical analysis, we demonstrate tight coupling between anonymization efficacy, business objectives, and data granularity, revealing that single-technique approaches fail to simultaneously ensure privacy and utility—especially under profit-driven enterprise constraints. We propose a synergistic anonymization framework integrating differential privacy, synthetic data generation, and other complementary strategies, with business semantics explicitly guiding technology selection and parameter tuning. Contribution/Results: The framework enhances interpretability and operational adaptability, enabling scalable, goal-oriented privacy engineering—a novel paradigm for practical, context-aware anonymization deployment.

Recent critiques have challenged the differential privacy guarantees of PATE-GAN and PrivBayes, questioning the validity of their privacy-utility trade-offs. However, these critiques rely on restrictive assumptions—such as synthetic or simplistic data distributions—and limited experimental settings, potentially biasing their conclusions. Method: We propose a more general privacy-utility evaluation framework that integrates privacy game analysis and theoretical verification, and conduct k-anonymity benchmarking experiments on real-world datasets without distributional assumptions. Contribution/Results: Under identical privacy budgets, both PATE-GAN and PrivBayes significantly outperform k-anonymity in statistical utility while maintaining strong differential privacy guarantees. We demonstrate that prior claims of “privacy failure” stem from flawed evaluation premises—specifically, the absence of rigorous privacy accounting and realistic data assumptions. Our empirical analysis refutes these criticisms and establishes synthetic data generation as a robust and effective privacy-enhancing technology.

This paper addresses the challenge of quantitatively assessing privacy risks—specifically identity and attribute disclosure—in practical synthetic data applications. We propose two interpretable, comparable, percentage-based metrics: RepU (Representation Uniqueness), measuring identity disclosure risk relative to original data, and DiSCO (Distributional Similarity Consistency), measuring attribute disclosure risk. We introduce the first well-defined, reproducible dual-metric evaluation framework, innovatively integrating key matching with target-variable consistency testing, augmented by statistical counting and adaptive thresholding. Crucially, the framework automatically filters out spurious disclosures attributable to domain-specific prior knowledge, substantially reducing false positives. Implemented as an extension to the R package *synthpop*, the method is empirically validated on real-world datasets, demonstrating robustness and practical utility for informed synthetic data release decisions.

Existing differential privacy (DP) synthetic data methods suffer substantial utility degradation when input data contain missing values. Method: This paper formally defines the DP synthetic data generation problem under missingness, characterizing how missingness mechanisms—Missing Completely at Random (MCAR), Missing at Random (MAR), and Missing Not at Random (MNAR)—propagate privacy loss and affect theoretical privacy bounds. Building on this analysis, we propose three adaptive synthesis strategies, each tailored to a specific missingness mechanism and privacy budget. Contribution/Results: Our key innovation is integrating missingness mechanism modeling directly into the DP synthetic data framework, enabling joint optimization of privacy protection and data utility. Evaluated on four real-world datasets with natural missingness, our approach achieves significant utility gains—improving downstream machine learning task accuracy by an average of 8.3%—while yielding tighter analytical privacy upper bounds compared to baseline methods.

Current evaluations of synthetic data anonymity are largely confined to the dataset level, overlooking the privacy risks inherent in the generative models themselves and thus failing to meet real-world regulatory compliance requirements. This work addresses this gap by adopting a model-centric perspective, establishing a novel privacy risk assessment framework that explicitly links the GDPR notion of “identifiability” with privacy attacks targeting generative models. Through a comparative analysis of differential privacy (DP) and similarity-based privacy metrics (SBPMs), the study demonstrates that SBPMs are insufficient in mitigating identifiability risks, whereas DP provides stronger, more reliable guarantees. The findings offer a more robust and responsible technical foundation for evaluating privacy and ensuring regulatory compliance in synthetic data systems.

This study addresses the challenge of sharing real-world educational data under strict privacy constraints, where existing differentially private synthetic data methods—often reliant on deep learning—are hindered by engineering complexity and limited practicality in small-sample, high-dimensional settings. The authors propose a training-free, two-stage framework: first, leveraging large language models to generate differentially private synthetic data for broad sharing and exploratory analysis; second, enabling on-demand validation of research findings on the original data through a secure remote code submission mechanism. Evaluated on three years of real educational data, the approach achieves synthetic data quality comparable to deep learning baselines while substantially reducing implementation overhead. Case studies show that approximately 36% of findings are reproducible on the真实 data, with the validation process introducing only negligible additional privacy loss.

Facing data exhaustion in model training and regulatory constraints on private data usage, this paper proposes the Generative Data Refinement (GDR) framework—the first to leverage pretrained generative models for *conditional synthetic data transformation*. GDR simultaneously achieves web-scale data detoxification, anonymization, and distribution preservation without manual prompt engineering. It safely converts private datasets containing sensitive or harmful content into high-quality, privacy-compliant training corpora while preserving semantic diversity and statistical fidelity. Experimental results demonstrate that GDR significantly outperforms state-of-the-art industrial baselines on both anonymization and detoxification tasks: it successfully purifies high-risk datasets and improves downstream model performance across multiple benchmarks. By enabling scalable, privacy-preserving data curation, GDR establishes a new paradigm for sustainable large language model training—one that reconciles dataset expansion with stringent privacy and safety requirements.

This work addresses the challenge of limited cross-institutional sharing of highly sensitive genomic data, which severely hinders the application of AI in data-scarce domains such as rare diseases. To overcome this barrier, the study proposes a novel framework that synergistically integrates secure multi-party computation (MPC) and differential privacy (DP). This approach simultaneously preserves input privacy—by safeguarding participants’ raw genomic data—and controls output privacy—by bounding information leakage from the synthesized data—enabling high-fidelity generative models to be trained collaboratively across institutions. Experiments on multiple real-world RNA-seq cohorts demonstrate that the resulting synthetic datasets achieve both strong privacy guarantees and high utility, effectively breaking down data silos while maintaining analytical validity.

Existing open-source tools lack a unified infrastructure, making it challenging to efficiently support the diverse modeling choices and privacy requirements inherent in synthetic data generation. This work proposes tidysynthesis, a modular and extensible meta-package that integrates a wide range of statistical modeling and differential privacy techniques through a unified declarative API. For the first time, it enables cross-framework algorithm composition and flexible customization of synthetic data workflows. The system substantially enhances both development efficiency and privacy guarantees, with its completeness, usability, and capacity to handle complex synthesis tasks demonstrated through empirical validation on U.S. Census survey data.