mlsanityData to 7 Oct 2026

Private GANs, Revisited

arXiv 2302.02936 ↗First posted 6 Feb 2023cs.LG cs.CR cs.CVcited by 17

Not on the map. Why?

Abstract (arXiv)

We show that the canonical approach for training differentially private GANs -- updating the discriminator with differentially private stochastic gradient descent (DPSGD) -- can yield significantly improved results after modifications to training. Specifically, we propose that existing instantiations of this approach neglect to consider how adding noise only to discriminator updates inhibits discriminator training, disrupting the balance between the generator and discriminator necessary for successful GAN training. We show that a simple fix -- taking more discriminator steps between generator steps -- restores parity between the generator and discriminator and improves results. Additionally, with the goal of restoring parity, we experiment with other modifications -- namely, large batch sizes and adaptive discriminator update frequency -- to improve discriminator training and see further improvements in generation quality. Our results demonstrate that on standard image synthesis benchmarks, DPSGD outperforms all alternative GAN privatization schemes. Code: https://github.com/alexbie98/dpgan-revisit.

Key facts. Cited by 17 papers in the map (data through 7 Oct 2026). Builds on 40 of its 66 references that are in the map. First posted on arXiv 6 Feb 2023.

Builds on

40 of 66 references.

References · 40

  1. 2014Adam: A Method for Stochastic Optimizationcited by 36,960
  2. 2019PyTorch: An Imperative Style, High-Performance Deep Learning Librarycited by 20,160
  3. 2018A Style-Based Generator Architecture for Generative Adversarial Networkscited by 5,986
  4. 2017Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithmscited by 4,283
  5. 2014Deep Learning Face Attributes in the Wildcited by 3,679
  6. 2019Analyzing and Improving the Image Quality of StyleGANcited by 3,580
  7. 2016Deep Learning with Differential Privacycited by 2,682
  8. 2018Large Scale GAN Training for High Fidelity Natural Image Synthesiscited by 2,609
  9. 2015Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networkscited by 2,048
  10. 2018Spectral Normalization for Generative Adversarial Networkscited by 1,618
  11. 2021Alias-Free Generative Adversarial Networkscited by 1,181
  12. 2022StyleGAN-XL: Scaling StyleGAN to Large Diverse Datasetscited by 489
28 more
  1. 2021Large Language Models Can Be Strong Differentially Private Learnerscited by 405
  2. 2021Differentially Private Fine-tuning of Language Modelscited by 392
  3. 2021Opacus: User-Friendly Differential Privacy Library in PyTorchcited by 375
  4. 2016Semi-supervised Knowledge Transfer for Deep Learning from Private Training Datacited by 347
  5. 2018Differentially Private Generative Adversarial Networkcited by 235
  6. 2018Subsampled R'enyi Differential Privacy and Analytical Moments Accountantcited by 234
  7. 2020Differentially Private Learning Needs Better Features (or Much More Data)cited by 233
  8. 2022Unlocking High-Accuracy Differentially Private Image Classification through Scalecited by 218
  9. 2019R'enyi Differential Privacy of the Sampled Gaussian Mechanismcited by 199
  10. 2021Numerical Composition of Differential Privacycited by 174
  11. 2020DP-CGAN: Differentially Private Synthetic Data and Label Generationcited by 118
  12. 2022Differentially Private Diffusion Modelscited by 118
  13. 2020GS-WGAN: A Gradient-Sanitized Approach for Learning Differentially Private Generatorscited by 117
  14. 2021Large-Scale Differentially Private BERTcited by 117
  15. 2021Benchmarking Differentially Private Synthetic Data Generation Algorithmscited by 80
  16. 2020DP-MERF: Differentially Private Mean Embeddings with Random Features for Practical Privacy-Preserving Data Generationcited by 78
  17. 2018Private Selection from Private Candidatescited by 75
  18. 2019Graphical-model based estimation and inference for differential privacycited by 73
  19. 2021Don't Generate Me: Training Differentially Private Generative Models with Sinkhorn Divergencecited by 54
  20. 2019G-PATE: Scalable Differentially Private Data Generator via Private Aggregation of Teacher Discriminatorscited by 53
  21. 2022Private Set Generation with Discriminative Informationcited by 34
  22. 2018Differentially Private Data Generative Modelscited by 33
  23. 2021DataLens: Scalable Privacy Preserving Training via Gradient Compression and Aggregationcited by 31
  24. 2021The Role of Adaptive Optimizers for Honest Private Hyperparameter Selectioncited by 25
  25. 2021Hermite Polynomial Features for Private Data Generationcited by 21
  26. 2020Private Post-GAN Boostingcited by 20
  27. 2020Taming GANs with Lookahead-Minmaxcited by 16
  28. 2021Differentially Private Normalizing Flows for Privacy-Preserving Density Estimationcited by 10

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2023: 6; 2024: 4; 2025: 4; 2026: 3.

17 papers.

Sort:

By area:

  1. 2023Differentially Private Diffusion Models Generate Useful Synthetic Imagescited by 85
  2. 2023DP-LDMs: Differentially Private Latent Diffusion Modelscited by 28
  3. 2023A Unified View of Differentially Private Deep Generative Modelingcited by 4
  4. 2024PFGuard: A Generative Framework with Privacy and Fairness Safeguardscited by 3
  5. 2024Differentially Private Synthetic Data Generation for Relational Databasescited by 2
  6. 2024Privacy without Noisy Gradients: Slicing Mechanism for Generative Model Trainingcited by 2
  7. 2025Privacy Amplification Through Synthetic Data: Insights from Linear Regressioncited by 2
  8. 2023Training generative models from privatized datacited by 1
  9. 2023Simulation-based Bayesian Inference from Privacy Protected Datacited by 1
  10. 2024Leveraging Programmatically Generated Synthetic Data for Differentially Private Diffusion Trainingcited by 1
  11. 2025Optimal Differentially Private Sampling of Unbounded Gaussianscited by 1
  12. 2025SoK: Can Synthetic Images Replace Real Data? A Survey of Utility and Privacy of Synthetic Image Generationcited by 1
5 more
  1. 2023DPAF: Image Synthesis via Differentially Private Aggregation in Forward Phasecited by 0
  2. 2025GenEAva: Generating Cartoon Avatars with Fine-Grained Facial Expressions from Realistic Diffusion-based Facescited by 0
  3. 2026Rethinking Benchmarks for Differentially Private Image Classificationcited by 0
  4. 2026Privacy Amplification Persists under Unlimited Synthetic Data Releasecited by 0
  5. 2026Decomposing Private Image Generation via Coarse-to-Fine Wavelet Modelingcited by 0