FETA-Pro Boosts Fidelity of Differentially Private Synthetic Images

Global: FETA-Pro Boosts Fidelity of Differentially Private Synthetic Images

A team of computer scientists has unveiled a new technique called FETA-Pro that enhances the quality of differentially private (DP) synthetic images. The method, detailed in a recent arXiv preprint, integrates frequency-based training shortcuts with traditional spatial features to improve DP training outcomes, particularly on datasets with diverse image content.

Background

Previous efforts to improve DP synthetic image generation have largely centered on refining core optimization algorithms such as DP‑Stochastic Gradient Descent (DP‑SGD). One notable approach, DP‑FETA, employed “central images” to warm up DP training before applying DP‑SGD, achieving gains primarily on homogenous datasets.

Introducing Frequency Features

According to the arXiv paper, the researchers observed that central‑image warm‑ups falter when image samples vary significantly. To address this gap, FETA‑Pro introduces frequency features as intermediate training shortcuts. These features capture spectral information that sits between the coarse spatial cues of central images and the full detail of raw images, enabling a more nuanced curriculum for DP training.

Model Architecture

The proposed pipeline leverages generative models in a staged fashion. An auxiliary generator first produces images aligned with noisy frequency features. A second model then consumes these generated images alongside spatial features while performing DP‑SGD. This modular design mitigates the training discrepancy that arises when combining disparate feature types.

Performance Gains

Evaluated across five sensitive image datasets, FETA‑Pro demonstrated an average of 25.7% higher fidelity and a 4.1% increase in utility compared with the best‑performing baseline, all under a privacy budget of ε = 1. These metrics indicate a notable improvement in both visual quality and downstream task performance while preserving strict differential privacy guarantees.

Future Directions

The authors suggest that the frequency‑feature shortcut could be extended to other modalities and that further exploration of multi‑model pipelines may yield additional privacy‑preserving benefits. By offering a flexible framework that accommodates diverse data characteristics, FETA‑Pro may influence future research on DP synthetic data generation.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.