Probabilistic Generative Models for Synthesizing Privacy-Preserving Big Data with Statistical Fidelity Guarantees

Abstract

The increasing demand for large-scale data sharing in data-driven research and industry has intensified concerns surrounding individual privacy and data confidentiality. Conventional privacy-preserving techniques such as anonymization, suppression, and heuristic perturbation have proven insufficient, particularly for high-dimensional big data, where linkage and inference attacks remain feasible. Synthetic data generation has therefore emerged as a promising alternative, enabling data dissemination while reducing direct exposure of sensitive records. Nonetheless, achieving rigorous privacy guarantees without sacrificing statistical fidelity and analytical utility remains a fundamental challenge.
This paper investigates probabilistic generative models as a principled solution for synthesizing privacy-preserving big data with formal guarantees. A unified framework is presented that integrates probabilistic generative modeling with differential privacy mechanisms to provide quantifiable protection against information leakage. The study examines Bayesian networks, variational autoencoders, and generative adversarial networks, incorporating advanced privacy accounting techniques such as Rényi differential privacy and moments-based analysis. Privacy budgets are carefully allocated, and noise is calibrated to data sensitivity during model training to balance privacy and utility.
Comprehensive experiments are conducted on benchmark tabular datasets to evaluate privacy protection, statistical fidelity, and downstream task performance. Results show that differentially private probabilistic generative models can preserve marginal distributions, correlation structures, and predictive accuracy under strict privacy constraints. Moreover, the generated synthetic datasets demonstrate strong resistance to membership inference attacks, indicating robustness against common adversarial threats. Overall, this work provides a systematic and empirically grounded foundation for trustworthy synthetic data generation, offering practical guidance for secure data sharing in sensitive domains and governance contexts.