The Distributed Discrete Gaussian Mechanism for Federated Learning with Secure Aggregation

Peter Kairouz, Ziyu Liu, Thomas Steinke
Proceedings of the 38th International Conference on Machine Learning, PMLR 139:5201-5212, 2021.

Abstract

We consider training models on private data that are distributed across user devices. To ensure privacy, we add on-device noise and use secure aggregation so that only the noisy sum is revealed to the server. We present a comprehensive end-to-end system, which appropriately discretizes the data and adds discrete Gaussian noise before performing secure aggregation. We provide a novel privacy analysis for sums of discrete Gaussians and carefully analyze the effects of data quantization and modular summation arithmetic. Our theoretical guarantees highlight the complex tension between communication, privacy, and accuracy. Our extensive experimental results demonstrate that our solution is essentially able to match the accuracy to central differential privacy with less than 16 bits of precision per value.

Cite this Paper

BibTeX
@InProceedings{pmlr-v139-kairouz21a, title = {The Distributed Discrete Gaussian Mechanism for Federated Learning with Secure Aggregation}, author = {Kairouz, Peter and Liu, Ziyu and Steinke, Thomas}, booktitle = {Proceedings of the 38th International Conference on Machine Learning}, pages = {5201--5212}, year = {2021}, editor = {Meila, Marina and Zhang, Tong}, volume = {139}, series = {Proceedings of Machine Learning Research}, month = {18--24 Jul}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v139/kairouz21a/kairouz21a.pdf}, url = {https://proceedings.mlr.press/v139/kairouz21a.html}, abstract = {We consider training models on private data that are distributed across user devices. To ensure privacy, we add on-device noise and use secure aggregation so that only the noisy sum is revealed to the server. We present a comprehensive end-to-end system, which appropriately discretizes the data and adds discrete Gaussian noise before performing secure aggregation. We provide a novel privacy analysis for sums of discrete Gaussians and carefully analyze the effects of data quantization and modular summation arithmetic. Our theoretical guarantees highlight the complex tension between communication, privacy, and accuracy. Our extensive experimental results demonstrate that our solution is essentially able to match the accuracy to central differential privacy with less than 16 bits of precision per value.} }
Endnote
%0 Conference Paper %T The Distributed Discrete Gaussian Mechanism for Federated Learning with Secure Aggregation %A Peter Kairouz %A Ziyu Liu %A Thomas Steinke %B Proceedings of the 38th International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2021 %E Marina Meila %E Tong Zhang %F pmlr-v139-kairouz21a %I PMLR %P 5201--5212 %U https://proceedings.mlr.press/v139/kairouz21a.html %V 139 %X We consider training models on private data that are distributed across user devices. To ensure privacy, we add on-device noise and use secure aggregation so that only the noisy sum is revealed to the server. We present a comprehensive end-to-end system, which appropriately discretizes the data and adds discrete Gaussian noise before performing secure aggregation. We provide a novel privacy analysis for sums of discrete Gaussians and carefully analyze the effects of data quantization and modular summation arithmetic. Our theoretical guarantees highlight the complex tension between communication, privacy, and accuracy. Our extensive experimental results demonstrate that our solution is essentially able to match the accuracy to central differential privacy with less than 16 bits of precision per value.
APA
Kairouz, P., Liu, Z. & Steinke, T.. (2021). The Distributed Discrete Gaussian Mechanism for Federated Learning with Secure Aggregation. Proceedings of the 38th International Conference on Machine Learning, in Proceedings of Machine Learning Research 139:5201-5212 Available from https://proceedings.mlr.press/v139/kairouz21a.html.

Related Material