A Unified Benchmark for Differentially Private Image Synthesis

View the PDF file for the paper entitled DPIMAGEBENCH: A unified standard for synthesis of differential photos, by Chen Gong and 3 other authors

PDF HTML (experimental) view

a summary:DEP (DP) aims to create artificial images that maintain sensitive image features while protecting the privacy of individual images within the data collection. Despite recent developments, we find that unintended evaluation protocols-and sometimes they were applied through studies. This not only hinders the understanding of the current methods, but also hinders future developments.

To address the problem, this dpimagebench paper is offered to synthesize DP image, with a studied design in several dimensions: (1) ways. We study eleven and a half -ways systematically, each on the basis of the structure of the model, the pre -training strategy, and the privacy mechanism. (2) Evaluation. We include nine collections of data and seven measures of sincerity and interest to accurately evaluate them. It is worth noting, we find that a common practice for choosing rated works on the basis of the highest accuracy in the sensitive test set not only violates the DP, but also works to estimate the degree of utility. DPIMAGEBENCH Correction of these errors. (3) platform. Despite the methods and evaluation protocols, DPIMAGEBENCH provides a unified interface that accommodates current and future applications within a unified framework. With DPIMAGEBENCH, we have many observed results. For example, unlike the common wisdom that usually training on public image data collections is useful, we find that the distribution similarity between pre -images and allergies greatly affects the performance of artificial images and does not always result in improvements. In addition, adding noise to low -dimensional features, such as the high -level properties of sensitive images, are less affected by the privacy budget compared to adding noise to high -dimensional features, such as weight gradients. Previous methods work better than the latter under a low private budget.