DGGI: Deep Generative Gradient Inversion with diffusion model

Liwen Wu, Zhizhi Liu, Bin Pu, Kang Wei, Hangcheng Cao, Shaowen Yao

Abstract: Highlights•An effective gradient inversion attack with Diffusion Models is proposed.•The proposed group consistency regularization can address the issue of spatial variations.•Diffusion models as prior knowledge is employed for effective image reconstruction.•Our method achieves favorable performance at larger batch sizes or under various defenses on natural and medical image datasets.