Abstract
We propose Gradient Inversion Transcript (GIT), a novel generative approachfor reconstructing training data from leaked gradients. GIT employs agenerative attack model, whose architecture is tailored to align with thestructure of the leaked model based on theoretical analysis. Once trainedoffline, GIT can be deployed efficiently and only relies on the leakedgradients to reconstruct the input data, rendering it applicable under variousdistributed learning environments. When used as a prior for other iterativeoptimization-based methods, GIT not only accelerates convergence but alsoenhances the overall reconstruction quality. GIT consistently outperformsexisting methods across multiple datasets and demonstrates strong robustnessunder challenging conditions, including inaccurate gradients, data distributionshifts and discrepancies in model parameters.