A Privacy-Preserving Framework for Generative Model-driven Synthetic Datasets

Debalina PadariyaAAAI 2025 · DOI

TL;DR

This paper proposes a novel framework for generating synthetic datasets using generative models (GMs) that prioritizes privacy. It addresses the challenge of balancing data utility with robust privacy guarantees, particularly for time-series data, by integrating differential privacy and developing optimized metrics for evaluating this trade-off.

What Problem Does This Paper Address?

The increasing use of generative models (GMs) like GANs and VAEs for creating synthetic data offers a promising solution for privacy preservation. However, these models can inadvertently leak sensitive information about the original training data, making them vulnerable to privacy attacks. Existing privacy-preserving methods, while improving, still face challenges in terms of accuracy, optimal privacy budget selection, and applicability across different domains. Furthermore, the reliability of synthetic data generation (SDG) is under scrutiny due to regulations like GDPR and CPRA. This paper aims to bridge this gap by developing a comprehensive framework that enhances privacy guarantees for GMs, especially in the underexplored domain of time-series data, while also proposing better ways to measure the effectiveness of these privacy-preserving techniques.

Methods

The research outlines a multi-phase approach. The initial phase involved a systematic literature survey to map the landscape of privacy-preserving GMs and to develop taxonomies for privacy and utility metrics. The second phase focused on developing a methodology for privacy-preserving synthetic data generation specifically for time-series applications. This involved creating an extensible framework that integrates generative models, attack modules, and defense mechanisms, with a strong emphasis on state-of-the-art differential privacy (DP) guarantees. The framework utilizes DPSGD (Differentially Private Stochastic Gradient Descent) for training models with DP. The future research plan involves integrating suitable privacy and utility metrics, proposing optimized metric suites for evaluating privacy/utility trade-offs, and developing a user-friendly library for practitioners.

Datasets

Not explicitly discussed in the paper. The paper focuses on the framework and methodology for generating synthetic data, particularly for time-series applications, rather than evaluating on specific public datasets.

Key Insights and Ideas

  • Time-series data is an underexplored frontier: The sequential nature and high dimensionality of time-series data present unique challenges for privacy-preserving synthetic data generation, making it a critical area for further research.
  • Differential Privacy is key, but needs careful calibration: While DP offers strong theoretical privacy guarantees, its practical implementation in GMs requires careful consideration of noise calibration to avoid compromising data utility.
  • Beyond individual metrics: The need for optimized suites: Evaluating the complex interplay between privacy and utility requires more than just individual metrics. Optimized metric suites, which intelligently compose and weight different metrics, are crucial for a holistic assessment.
  • Bridging the gap for practitioners: A practical, usable framework with a user-facing interface is essential to enable wider adoption of privacy-preserving synthetic data generation in industry and the public sector.

Key Results

The paper describes the development of a methodology and an extensible framework for privacy-preserving synthetic data generation for time-series applications. An extensive evaluation demonstrated the framework's effectiveness in balancing privacy and utility across different privacy budgets, showing competitive performance against state-of-the-art models. The research has also led to the design of novel taxonomies for categorizing privacy and utility metrics.

Limitations and Future Work

The paper acknowledges that while DP offers rigorous statistical assurance, careful consideration is needed when calibrating the noise. Future work includes:

  • Integrating and optimizing metric suites: Developing and testing optimized metric suites to effectively evaluate the privacy/utility trade-offs in synthetic data.
  • Evaluating privacy/utility trade-offs: Conducting a thorough evaluation of these trade-offs across different synthetic data generation configurations.
  • Deployment: Creating a usable library that integrates the framework and provides a user-facing interface for practitioners to perform privacy risk assessments on their own datasets.

Summary updated

BibTeX
@inproceedings{padariya2025privacy,
  author = {Debalina Padariya},
  title = {{A Privacy-Preserving Framework for Generative Model-driven Synthetic Datasets}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i28.35222},
}
Figure