We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:


Current browse context:


Change to browse by:

References & Citations

DBLP - CS Bibliography


(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Machine Learning

Title: PRECODE - A Generic Model Extension to Prevent Deep Gradient Leakage

Abstract: Collaborative training of neural networks leverages distributed data by exchanging gradient information between different clients. Although training data entirely resides with the clients, recent work shows that training data can be reconstructed from such exchanged gradient information. To enhance privacy, gradient perturbation techniques have been proposed. However, they come at the cost of reduced model performance, increased convergence time, or increased data demand. In this paper, we introduce PRECODE, a PRivacy EnhanCing mODulE that can be used as generic extension for arbitrary model architectures. We propose a simple yet effective realization of PRECODE using variational modeling. The stochastic sampling induced by variational modeling effectively prevents privacy leakage from gradients and in turn preserves privacy of data owners. We evaluate PRECODE using state of the art gradient inversion attacks on two different model architectures trained on three datasets. In contrast to commonly used defense mechanisms, we find that our proposed modification consistently reduces the attack success rate to 0% while having almost no negative impact on model training and final performance. As a result, PRECODE reveals a promising path towards privacy enhancing model extensions.
Comments: 16 pages, 16 figures, 5 tables (supplementary material included). Accepted at WACV 2022. Code available at this https URL
Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
Cite as: arXiv:2108.04725 [cs.LG]
  (or arXiv:2108.04725v2 [cs.LG] for this version)

Submission history

From: Daniel Scheliga [view email]
[v1] Tue, 10 Aug 2021 14:43:17 GMT (6447kb,D)
[v2] Thu, 21 Oct 2021 08:44:37 GMT (6447kb,D)

Link back to: arXiv, form interface, contact.