Reproduce “Tabular Data Synthesis with Differential Privacy: A Survey” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Mengmeng Yang, Chi-Hung Chi, Kwok-Yan Lam, Jie Feng, Taolin Guo, Wei Ni

Researchers from CSIRO Data61, Nanyang Technological University, Xidian University, and Chongqing Normal University compiled a comprehensive survey of differentially private tabular data synthesis, classifying methods by statistical or deep learning approaches in both centralized and distributed settings. The work analyzes evaluation techniques for fidelity, utility, and privacy, and identifies seven key research gaps for future development in the field.

Key insights

  • Differentially private tabular data synthesis methods vary significantly in their ability to capture data dependencies, handle mixed data types, and manage privacy-utility trade-offs across centralized and distributed environments.
  • Deep learning models, especially GANs, require specific adaptations for tabular data to overcome challenges like mode collapse and effectively preserve complex correlations and mixed data types.
  • Evaluation of synthetic data necessitates a holistic approach, considering metrics for fidelity, utility, and privacy simultaneously to provide a comprehensive assessment of data quality and security.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.