Skip to content

Latest commit

 

History

History
111 lines (81 loc) · 4.47 KB

readme.md

File metadata and controls

111 lines (81 loc) · 4.47 KB
Raw

DPLab: Benchmarking Differential Privacy Aggregation Operations

This repo targets to provide a unified interface to access and evaluate the same aggregation functionalities in different open-source differential privacy (DP) libraries. With a simple CLI, one can choose the library, the aggregation function, and many other experimental parameters and apply the specified DP measurement to data stored in a .csv file. The repo also provides both synthetic and real-world example datasets for evaluation purposes. Evaluation results are stored in a .json file and metrics are provided for repeated experiments. The repo also provides a CLI tool to generate configuration groups for larger-scale comparison experiments.

1-Min Tutorial

Get hands-on in 1 minute with our tutorial notebook.

dplab_architecture

Currently supported aggregation operations:

  • COUNT
  • SUM
  • MEAN
  • VAR
  • MEDIAN
  • QUANTILE

Currently supported libraries:

Installation

To install dplab, one can use the package on pypi

pip install dplab

Or with source code: clone the repo, switch the working directory, and install the dependencies

git clone [email protected]:camelop/dp_lab.git
cd dp-lab
pip install -e .

To use tmlt

export PYSPARK_PYTHON=/usr/bin/python3
sudo apt install openjdk-8-jre-headless
pip3 install -i https://d3p0voevd56kj6.cloudfront.net python-flint
pip3 install tmlt.analytics

To use chorus, please make sure you have Java runtime installed. (If you have already installed tmlt, it should be fine.)

How to run dp libraries in the benchmark

Run a specific library with the CLI

dplab_run <library> <operation> <input_file> <output_file> <other options>

For example:

dplab_run pydp sum data/1.csv data/1.json -f -r 1000

Other options include:

  • mode: Evaluation mode, one can choose from "plain" (no timing/mem measurement), "internal" (internal measurement), or "external" (external tracking).
  • epsilon: DP parameter, default is set to 1.
  • quant: Quantile value for QUANTILE operation, a float number between 0 and 1.
  • lb: The optional value lower bound estimation used when applying certain differential privacy aggregations.
  • ub: The optional value upper bound estimation used when applying certain differential privacy aggregations.
  • repeat: How many time should the evaluation repeat.
  • force: Force to overwrite the output file.
  • debug: Include debugging information in the output file.
  • python_command: Python command used to run the script in the external mode.
  • external_sample_interval: timing/mem consumption sample interval in the external mode.

For more information, please check the main entry file.

Generating synthetic data

# Make sure you are in the root directory of the repo
# Data will be generated in the ./data/ directory
# The procedure will generate about 28GB of data
# To avoid the risk of running out of disk space, you can comment out the performance test lines (Line26-27) in SYN_TARGETS defined in the script
python3 scripts/gen_data.py

How to run experiments in the benchmark

Generate the experiment commands, this will generate an ./exp.db.json file under the working directory (you can also use --location to specify a different place).

dplab_exp plan --repeat 100 --group_num 100

Queue the experiments for execution

dplab_exp launch --debug

The command updates the results to exp.db.json.

One can potentially view the results via

python3 scripts/view_exp_db.py