This repository contains the artifact for the IEEE S&P 2024 paper:
N. Küchler, E. Opel, H. Lycklama, A. Viand, and A. Hithnawi,
"Cohere: Managing Differential Privacy in Large Scale Systems"
in 2024 IEEE Symposium on Security and Privacy (SP), San Francisco, CA, USA, 2024
Bibtex
@INPROCEEDINGS {Kuchler2024-dpcohere,
author = {Küchler, Nicolas and Opel, Emanuel and Lycklama, Hidde and Viand, Alexander and Hithnawi, Anwar},
booktitle = {2024 IEEE Symposium on Security and Privacy (SP)},
title = {Cohere: Managing Differential Privacy in Large Scale Systems},
year = {2024},
volume = {},
issn = {2375-1207},
pages = {122-122},
url = {https://doi.ieeecomputersociety.org/10.1109/SP54263.2024.00122},
publisher = {IEEE Computer Society},
address = {Los Alamitos, CA, USA},
month = {may}
}
The existing implementation of Cohere serves as an academic prototype primarily focusing on the DP resource planner. This component is the core of Cohere's design and effectively handles most of the complexity of DP management. In addition, we provide:
- A hyperparameter explorer for investigating parameter tradeoffs and assessing the suitability of Cohere in different scenarios.
- A customizable workload generator enabling the creation of complex mixed workloads of DP applications.
- A request adapter showcasing the generation of Cohere requests directly from Tumult Analytics, a DP library designed for aggregate queries on tabular data, as well as Opacus, a PyTorch-based DP library for ML training.
The evaluation of Cohere is orchestrated with the doe-suite experiment management tool, with configuration details available in doe-suite-config.
Get started by installing and running the DP resource planner on your local machine. For this, we include a basic workload in the repository. If you need more complex workloads, you can use the workload generator.
To reproduce the paper's results on AWS or a SLURM-based scientific compute cluster, check out the documentation in doe-suite-config.