GS-WGAN: A Gradient-Sanitized Approach for Learning Differentially Private Generators

Authors: Dingfan Chen, Tribhuvanesh Orekondy, Mario Fritz

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through extensive experiments, we find our approach consistently outperforms state-of-the-art approaches across multiple metrics (e.g., sample quality) and datasets. Extensi ve evaluations on various datasets demonstrate that our method significantly improves the sample quality of privacy-preserving data over state-of-the-art approaches. To validate the applicability of our method to high-dimensional data, we conduct experiments on image datasets.
Researcher Affiliation Academia 1 CISPA Helmholtz Center for Information Security 2 Max Planck Institute for Informatics
Pseudocode No The paper does not contain explicitly labeled pseudocode or algorithm blocks. It describes the methods in text and uses mathematical formulas.
Open Source Code No The paper mentions using source code for baseline methods but does not state that the code for its own methodology is being released or provide a link to it.
Open Datasets Yes In line with previous works, we use MNIST [23] and Fashion-MNIST [40] dataset. We conduct experiments on the Federated EMNIST dataset [9] and compare our GS-WGAN with Fed-Avg GAN [3].
Dataset Splits No The paper mentions using 60k privately-generated data points for downstream tasks and evaluating during training iterations, but it does not specify explicit train/validation/test splits (e.g., percentages or counts) for the original datasets used in their experiments.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., GPU/CPU models, memory) used for running the experiments.
Software Dependencies No The paper mentions 'scikit-learn [34]' but does not provide a version number for it or any other software dependencies crucial for replication.
Experiment Setup Yes Better model architecture: While previous works are limited to shallow networks and thereby bottle-necking generated sample quality, our framework allows stable training with a complex model architecture (DCGAN [35] architecture for the discriminator, Res Net architecture (adapted from Big GAN [8]) for the generator) to help improve the sample quality; Discriminator warm-starting: To bootstrap the training process, we pre-train discriminators along with a non-private generator for a few steps, and we subsequently train the private generator using the warm-starting values of the discriminators. We evaluate the sample quality of our method considering multiple choices of subsampling rates (γ [1/250, 1/500, 1/1000, 1/1500]) over the training iterations. For a fair comparison, we evaluate all methods with a privacy budget of (ε, δ)=(10, 10 5) (consistently used in previous works) over 60K generated samples.