Towards High Quality DP Image Synthesis 117 views

Machine-learning models trained on sensitive image data may disclose information about their training records. Differentially private (DP) data synthesis addresses this risk by releasing artificial datasets through a mechanism with a formal DP guarantee. Although DP synthesis has achieved useful results for lower-dimensional data, high-dimensional image synthesis remains difficult because clipping and noise can substantially reduce the fidelity and downstream utility of learned generative models. This dissertation studies record-level DP image synthesis from four complementary directions. First, direct optimization develops DP-FETA and FETA-Pro, which use low-sensitivity private summaries to warm up generative models before DP-SGD training. Second, data-driven guidance develops PrivImage, which privately estimates the semantic distribution of a sensitive dataset and uses the estimate to select a compact public pretraining subset. Third, model-based transfer develops DP-SAPF, which privately selects salient parameter matrices in a public diffusion model and fine-tunes only the selected low-rank adapters with DP-SGD. Fourth, standardized evaluation develops DPImageBench, a unified framework that evaluates twelve representative methods across nine datasets and seven fidelity and utility metrics. DPImageBench also introduces Noisy SenV, a private checkpoint-selection protocol that avoids selecting downstream classifiers on the final test set. The formal guarantees in this dissertation are primarily image-record-level guarantees: when one person or patient contributes multiple images, they do not by themselves imply person-level DP. Together, these four directions cover the full pipeline of DP image synthesis, from algorithm design, to data and model utilization, and finally evaluation, with the goal of building practical privacy-preserving generative models.

PHD (Doctor of Philosophy)

Differentially Private Image Synthesis; Data Privacy; High-Dimensional Data Synthesis

English

2026-07-13