Improving Private Training via In-distribution Public Data Synthesis and Generalization

Jinseong Park, Yujin Choi, Jaewook Lee

Keywords: Differential Privacy, Privacy, Optimization, DP-SGD, Diffusion, Synthesis

TL;DR: Enhancing differentially private classification through in-distribution public data using diffusion synthesis and optimization for well-generalizing minima.

Abstract: To alleviate the utility degradation of deep learning classification with differential privacy (DP), employing extra public data or pre-trained models has been widely explored. Recently, the use of in-distribution public data has been investigated, where a tiny subset of data owners share their data publicly. In this paper, to mitigate memorization and overfitting by the limited-sized in-distribution public data, we leverage recent diffusion models and employ various augmentation techniques for improving diversity. We then explore the optimization to discover flat minima to public data and suggest weight multiplicity to enhance the generalization of the private training. While assuming 4\% of training data as public, our method brings significant performance gain even without using pre-trained models, i.e., achieving 85.78\% on CIFAR-10 with a privacy budget of $\varepsilon=2$ and $\delta=10^{-5}$.

Primary Area: societal considerations including fairness, safety, privacy

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 2957