DGGI: Deep Generative Gradient Inversion with diffusion model

Abstract

Federated learning is a privacy-preserving distributed framework that facilitates information fusion and sharing among different clients, enabling the training of a global model without exposing raw data. However, the gradient inversion attack that can reconstruct the training data via gradients has posed a significant threat. Prior attack approaches have demonstrated the efficacy of gradient inversion on low-resolution datasets with small batch sizes, which is impractical in real scenarios. To tackle this issue, this paper proposes an innovative and practical gradient inversion method, namely Deep Generative Gradient Inversion (DGGI), which employs the prior knowledge of diffusion models to enhance reconstruction performance on high-resolution datasets and larger batch sizes. Furthermore, a novel group consistency regularization term that constrains the distance between reconstruction and alignment images has been developed to address the issue of spatial variations caused by pre-trained diffusion model. Experiments conducted on both natural and medical image datasets demonstrate that our DGGI method outperforms state-of-the-art baselines in image reconstruction metrics. Furthermore, our approach achieves pixel-level reconstruction and causes leakage of privacy information, even at larger batch sizes or under various defenses, which can aid in the exploration of latent security concerns within information fusion models.