Gradient Inversion Transcript: Leveraging Robust Generated Priors to Reconstruct Training Data from Gradient Leakage

ICLR 2026 Conference Submission17128 Authors

Keywords: gradient inversion, training data reconstruction, image prior

Abstract:

We propose Gradient Inversion Transcript (GIT), a novel model-based approach for reconstructing training data from leaked gradients. GIT employs a data reconstruction model, whose architecture is tailored to align with the inversion of the federated learning (FL) model's back-propagation process. Once trained offline, GIT can be deployed efficiently and only relies on the leaked gradients to reconstruct the input data, rendering it applicable under various distributed learning environments.
When used as a prior for other iterative optimization-based methods, GIT not only accelerates convergence but also enhances the overall reconstruction quality. GIT consistently outperforms existing methods across multiple datasets and demonstrates strong robustness under challenging conditions, including inaccurate gradients, data distribution shifts and discrepancies in model parameters.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 17128