Gradient Inversion Transcript: A Generative Model to Reconstruct Training Data by Gradient Leakage

Xinping Chen, Chen Liu

Keywords: distributed learning, training data reconstruction, generative model, gradient inversion

Abstract:

We propose Gradient Inversion Transcript (GIT), a generic approach for reconstructing training data from gradient leakage in distributed learning using a generative model. Unlike traditional gradient matching techniques, GIT requires only the model architecture information, without access to the model's parameters, making it more applicable to real-world distributed learning settings. Additionally, GIT operates offline, eliminating the need for intensive gradient requests and online optimization.
Compared to existing generative methods, GIT adaptively constructs a generative network, with an architecture specifically tailored to the structure of the distributed learning model. Our extensive experiments demonstrate that GIT significantly improves reconstruction accuracy, especially in the case of deep models.
In summary, we offer a more effective and theoretically grounded strategy for exploiting vulnerabilities of gradient leakage in distributed learning, advancing the understanding of privacy risks in collaborative learning environments.

Primary Area: generative models

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 6565