Don’t Generate Me: Training Differentially Private Generative Models with Sinkhorn Divergence

Authors: Tianshi Cao, Alex Bie, Arash Vahdat, Sanja Fidler, Karsten Kreis

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct experiments on differentially private conditional image synthesis, since our focus is on generating high-dimensional data with privacy protection. We evaluate our method on both visual quality and data utility for downstream classification tasks. Additional experiments and analyses of the proposed semi-debiased Sinkhorn loss can be found in the Appendix.
Researcher Affiliation Collaboration Tianshi Cao1,2,4 Alex Bie3 Arash Vahdat4 Sanja Fidler1,2,4 Karsten Kreis4 1University of Toronto 2Vector Institute 3University of Waterloo 4NVIDIA
Pseudocode Yes Algorithm 1 DP-Sinkhorn
Open Source Code No Project page: https://nv-tlabs.github.io/DP-Sinkhorn. Code will be released through the project page4.
Open Datasets Yes Datasets We use 3 image datasets: MNIST [45], Fashion-MNIST [46], and Celeb A [47] downsampled to 32x32 pixels.
Dataset Splits No The paper does not explicitly state training, validation, and test dataset splits with percentages or sample counts. It mentions evaluation on a "synthetic dataset" and "real test sets" but not the splits for the original datasets.
Hardware Specification Yes In our experiments, DP-Sinkhorn can fit comfortably on an 11GB GPU, while GS-WGAN requires 24GB of GPU memory.
Software Dependencies No Our models are implemented in Py Torch. The paper mentions PyTorch but does not provide a specific version number or details for other software dependencies.
Experiment Setup Yes We set λ=0.05 for MNIST and Fashion-MNIST experiments, and λ=5 for Celeb A experiments. Complete implementation details can be found in the Appendix.