Enhancing gradient inversion attacks in federated learning via hierarchical feature optimization

The researchers propose GIFD, a federated learning gradient inversion attack method that can reconstruct sensitive data from shared gradients. The method decomposes the GAN model, progressively advancing the optimization target from the initial latent space layer by layer to intermediate-layer features closer to the output image, and uses an l1-ball constraint regularization term to avoid generating unrealistic images. GIFD is also extended to OOD scenarios and uses a label mapping technique to handle label inconsistency; experiments show that it can achieve pixel-level reconstruction and outperforms competing baselines in various federated learning scenarios.