Privacy Blur: Quantifying the Privacy and Utility of Image Data Release

In the arXiv paper Privacy Blur (arXiv:2512.16086), the authors quantify the privacy and utility of image data release, pointing out that the Gaussian blur used in practice can be broken by inversion attacks under low-precision implementations and has the worst privacy among the four de-identification methods compared. The study evaluates privacy using inversion and discrimination attacks and evaluates utility using the quality of representations obtained by training models on de-identified face data, comparing three alternatives: pixelation, noised pixelation (DP-Pix), and cropping, and finds that pixelation and DP-Pix at appropriate granularity can simultaneously provide privacy and utility for multiple computer vision tasks. The authors provide the proposed method and recommended parameters as a software package, Privacy Blur.