Preprint · 2026 · Open access

DP-DCGAN: Differential Privacy-Deep Convolutional Generative Adversarial Networks with Adaptive Gradient Perturbation

Libo Zhu, Xiang Chen

Research Square · Published 9 July 2026 · DOI 10.21203/rs.3.rs-10050492/v1

Open PDFView at the publisherCite this paper

In brief

“DP-DCGAN: Differential Privacy-Deep Convolutional Generative Adversarial Networks with Adaptive Gradient Perturbation” is a 2026 preprint by Libo Zhu and Xiang Chen, published in Research Square. OpenAlex counts 0 citations as of 19 November 2026. It is open access under a CC BY licence: the full text is free to read at Research Square.

Citations
0
References
48
Authors
2

Abstract

As one of the most dynamic generative models, generative adversarial network (GAN) has shown excellent performance in image synthesis, speech generation, data augmentation and other tasks. However, its high fitting ability to the original data distribution also increases the risk of privacy leakage. Differential privacy (DP) provides a theoretical framework to limit the influence of a single sample in a statistical sense, which can give strict privacy protection for data publishing and model training. In order to balance data availability and privacy security, this paper proposes a deep convolutional generative adversarial networks (DCGAN) framework based on DP (DCGAN with DP). Firstly, the framework takes the classical DCGAN as the backbone and learns the latent distribution through the generator-discriminator adversarial structure. Secondly, in the gradient update stage of the discriminator, the norm is first trimmed, and then Gaussian noise satisfying \((\epsilon)\)-DP is injected to ensure the privacy of individual samples during the whole training process. Finally, the \((\epsilon)\)-adaptive cropping strategy and multi-discriminator evaluation are combined to achieve controllable and robust privacy-friendly image synthesis, taking into account both privacy budget and generation quality. Through experiments on standard datasets (MNIST and Fashion-MNIST), DCGAN with DP achieves the best overall results across the three metrics, obtaining the highest average accuracy of 0.797 on MNIST and 0.659 on Fashion-MNIST, the strongest MIA resistance of 0.942/0.923, and the lowest ASR of 0.163/0.140, demonstrating superior generation utility and privacy protection.Our work provides an efficient and feasible new paradigm for privacy-friendly data synthesis, which can be extended to highly privacy-sensitive and quality-sensitive scenarios.

Cite this paper

Formatted from the catalogue record. Check names, capitals and page numbers against the paper before you submit.

APA 7

Zhu, L., & Chen, X. (2026). DP-DCGAN: Differential Privacy-Deep Convolutional Generative Adversarial Networks with Adaptive Gradient Perturbation. Research Square. https://doi.org/10.21203/rs.3.rs-10050492/v1

MLA 9

Zhu, Libo, and Xiang Chen. “DP-DCGAN: Differential Privacy-Deep Convolutional Generative Adversarial Networks with Adaptive Gradient Perturbation.” Research Square, 2026. https://doi.org/10.21203/rs.3.rs-10050492/v1.

Chicago

Zhu, Libo and Xiang Chen. 2026. “DP-DCGAN: Differential Privacy-Deep Convolutional Generative Adversarial Networks with Adaptive Gradient Perturbation.” Research Square. https://doi.org/10.21203/rs.3.rs-10050492/v1.

Harvard

Zhu, L. and Chen, X. (2026) ‘DP-DCGAN: Differential Privacy-Deep Convolutional Generative Adversarial Networks with Adaptive Gradient Perturbation’, Research Square. doi:10.21203/rs.3.rs-10050492/v1.

IEEE

L. Zhu and X. Chen, “DP-DCGAN: Differential Privacy-Deep Convolutional Generative Adversarial Networks with Adaptive Gradient Perturbation,” Research Square, 2026, doi: 10.21203/rs.3.rs-10050492/v1.

BibTeX

@misc{zhu2026dcgan,
  title     = {{DP-DCGAN: Differential Privacy-Deep Convolutional Generative Adversarial Networks with Adaptive Gradient Perturbation}},
  author    = {Zhu, Libo and Chen, Xiang},
  howpublished = {Research Square},
  year      = {2026},
  doi       = {10.21203/rs.3.rs-10050492/v1},
  url       = {https://doi.org/10.21203/rs.3.rs-10050492/v1}
}

Download for a reference manager: BibTeX (.bib) · RIS for Zotero, Mendeley and EndNote (.ris)

Topic and keywords

Physical Sciences › Computer Science › Artificial Intelligence › Privacy-Preserving Technologies in Data

  • differential privacy
  • generative adversarial networks
  • DCGAN
  • gradient perturbation
  • privacy leakage
  • image generation
  • privacy-preserving data synthesis
  • membership inference attacks

Work with this paper

Most-cited papers in Computer Science

All highly cited Computer Science papers

Questions about this paper

How do I cite this paper in APA style?

Zhu, L., & Chen, X. (2026). DP-DCGAN: Differential Privacy-Deep Convolutional Generative Adversarial Networks with Adaptive Gradient Perturbation. Research Square. https://doi.org/10.21203/rs.3.rs-10050492/v1

How many times has this paper been cited?

0 times, according to OpenAlex on 19 November 2026. Google Scholar and Scopus count different sets of sources, so their numbers differ.

Is this paper free to read?

Yes. It is open access at Research Square: sign in and open the PDF from this page.

Who wrote this paper?

Libo Zhu, Xiang Chen.

What is the DOI of this paper?

10.21203/rs.3.rs-10050492/v1, permanently at https://doi.org/10.21203/rs.3.rs-10050492/v1. Sign in and open it from this page.

Bibliographic data from OpenAlex (CC0), last checked 19 November 2026. ResearchGuru does not host this paper; once you are signed in, the links open it at the publisher or the repository that does.