CGGL: A client-side generative gradient leakage attack with double diffusion prior

Open Webpage

Bin Pu, Zhizhi Liu, Liwen Wu, Kai Xu, Bocheng Liang, Ziyang He, Benteng Ma, Lei Zhao

Published: 01 Jan 2025, Last Modified: 06 Nov 2025Inf. Fusion 2025EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Highlights•A novel client-side gradient inversion attack with double diffusion prior is proposed.•Adaptive poisoning attack effectively mitigates the effects of gradient aggregation.•A double diffusion prior-based attack framework improves image reconstruction quality.•CGGL boosts the quality of reconstructed images and shows strong attack performance.