GIDM: A High-Fidelity Gradient Inversion Method Using Diffusion Model in Federated Learning

Gradient Inversion attacks pose a significant threat to Federated Learning (FL) by recovering users' private data from shared gradients. Recent studies have demonstrated the feasibility of reconstructing private data in image classification tasks within FL. However, these methods usually reconstruct images with limited fidelity or rely on unrealistic assumptions. In this paper, we propose a high-fidelity Gradient Inversion method using Diffusion Model (GIDM), which incorporates latent diffusion model into gradient inversion attack. Rather than using the diffusion model's image generation capability as a black box, we disassemble the model and treat each sampling step as a fundamental operating unit, optimizing the intermediate output. Additionally, we design a feature regularization to serve as an auxiliary part of the optimization objective. By reconstructing the high-dimensional features of truth images, GIDM uncovers additional private information, thereby enhancing the fidelity of the reconstructed images. Extensive experiments demonstrate that GIDM achieves pixel-level image reconstruction with superior fidelity and visual quality compared to existing GAN-based methods. Furthermore, GIDM proves effective against defensive strategies and large data batches, underscoring the need for more advanced privacy protection mechanisms in FL.