Differentially Private Normalizing Flows for Synthetic Tabular Data Generation

Authors: Jaewoo Lee, Minjung Kim, Yonghyun Jeong, Youngmin Ro7345-7353

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our empirical evaluations show that the proposed model preserves statistical properties of original dataset better than other baselines.
Researcher Affiliation Collaboration 1University of Georgia 2Samsung SDS jwlee@cs.uga.edu, {mj100.kim, yhyun.jeong, youngmin.ro}@samsung.com
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide concrete access to source code for the methodology described.
Open Datasets Yes To evaluate the performance of DP-HFlow, we perform experiments on four real datasets: Adult, Census, Covertype, Intrusion.
Dataset Splits No The paper mentions evaluating on a 'testing' set but does not provide specific details on training, validation, and test splits (e.g., percentages or counts) or a general splitting methodology.
Hardware Specification Yes All experiments were performed on a server with an NVIDIA RTX 8000 GPU.
Software Dependencies No The paper mentions "Ada Belief optimizer (Zhuang et al. 2020)" but does not specify software names with version numbers for replication.
Experiment Setup Yes In all experiments, DP-HFlow is instantiated by stacking 3 blocks of autoregressive spline transformation and low rank-based linear transformation on top of dequantization layer. A reverse ordering permutations is inserted in between blocks. We used Ada Belief optimizer (Zhuang et al. 2020) with learning rate 0.001 and default smoothing parameter of β1 = 0.9 and β2 = 0.999.