SDNist: datasets and evaluation tools for data synthesizers
Project description
SDNist: Benchmark Data and Evaluation Tools for Data Synthesizers
This package provides tools for standardized and reproducible comparison of synthetic generator models on real-world data and use cases. Both datasets and metrics were developed for and vetted through the NIST PSCR Differential Privacy Temporal Map Challenge.
Quick introduction
You have two possible workflows:
- Manually manage the public and private datasets as
pandas.DataFrame
objects, directly generate your synthetic data, and directly compute the score. - Reproduce the setup of the challenge---i.e., create a synthesizer subclass of
challenge.submission.Model
, then callrun(model, challenge="census")
. This makes sure your synthesizer is scored against the same datasets as in the Challenge.
In all cases, the scoring does not numerically check whether your synthesizer is actually $\epsilon$-differentially private or not. You have to provide a formal proof yourself.
Installation
Requirements: Python >=3.6
The SDNist source code is hosted on Github, and all the data tables should be downloaded from the SDNist Github Releases. Alternatively, the data can be manually downloaded as part of the latest release SDNist Release 1.3.0.
- Data Download Notes:
- SDNist does not download just a specific dataset; instead, it downloads all the available datasets in the library.
- If data is manually downloaded, copy the contents inside the 'data' directory from the extracted zip file to your data root directory.
- The default root data directory of SDNist is
<your-current-working-directory>/data
. The current working directory is the directory in which you run the SDNist library through a console/terminal, or the directory that contains your Python or IPython files that imports the SDNist library.
- Install
sdnist
Python module through git repository:
git clone https://github.com/usnistgov/SDNist && cd SDNist
pip install .
- Install
sdnist
Python module through git in a virtual environment:
MAC OS / Linux
git clone https://github.com/usnistgov/SDNist && cd SDNist
python3 -m venv venv
. venv/bin/activate
pip install .
Windows
git clone https://github.com/usnistgov/SDNist && cd SDNist
python3 -m venv venv
. venv/Scripts/activate
pip install .
Contributions
This repository is being actively developed, and we welcome contributions.
If you encounter a bug, please create an issue.
Feel free to create a Pull Request to help us correct bugs and other issues.
Please contact us if you wish to augment or expand existing features.
Examples
1) Option 1 (quickest)
Loading and scoring
>>> import sdnist
>>> dataset, schema = sdnist.census() # Retrieve public dataset
>>> dataset.head()
PUMA YEAR HHWT GQ ... POVERTY DEPARTS ARRIVES sim_individual_id
0 17-1001 2012 88.0 1 ... 118 902 909 12
1 17-1001 2012 61.0 1 ... 262 732 744 33
2 17-1001 2012 54.0 1 ... 118 642 654 401
3 17-1001 2012 106.0 1 ... 262 0 0 470
4 17-1001 2012 31.0 1 ... 501 0 0 702
[5 rows x 36 columns]
>>> synthetic_dataset = dataset.sample(n=20000) # Build a fake synthetic dataset
# Compute the score of the synthetic dataset
>>> sdnist.score(dataset, synthetic_dataset, schema, challenge="census")
100%|███████████████████████████████████████████| 50/50 [00:04<00:00, 12.11it/s]
CensusKMarginalScore(847)
Discretizing a dataset
Many synthesizers require working on categorical/discretized data, yet many features of sdnist
datasets are actually
integer or floating point valued. sdnist
provide a simple tool to discretize/undiscretize sdnist
datasets.
First, note that the k-marginal score itself works on categorical data under the hood. For fairness, the bins that are used can be considered public. They are available as follows:
The ACS (American Community Survey) dataset:
>>> bins = sdnist.kmarginal.CensusKMarginalScore.BINS
The Chicago taxi dataset:
>>> bins = sdnist.kmarginal.TaxiKmarginalScore.BINS
The pd.DataFrame
datasets can then be discretized using
>>> dataset_binned = sdnist.utils.discretize(dataset, schema, bins)
sdnist.utils.discretize
returns a pd.DataFrame
where each value is remapped to (0, n-1)
where n
is the number of distinct values. Note that the even though the score
functions should be given unbinned datasets (i.e., if your synthesizer works on discretized dataset), you should first undiscretize your synthetic data. This can be done using
>>> synthetic_dataset_binned = ... # generate your synthetic data using your own method
>>> synthetic_dataset = sdnist.utils.undo_discretize(synthetic_dataset_binned, schema, bins)
Directly computing the score on a given .csv
file
You can directly run from a terminal:
% python -m sdnist your_file.csv
This will score against the public census (ACS) dataset and display the results on an html page:
To score the synthetic dataset against one of the test datasets:
% python -m sdnist your_synthetic_ga_nc_sc.csv --test-dataset GA_NC_SC_10Y_PUMS
Other options are available by calling --help
.
Computing aggregate score for all synthetic files generated using different epsilon values
To generate a final aggregate score over all epsilon values for the Census Challenge:
% python -m sdnist.challenge.submission
To score synthetic data file for dataset GA_NC_SC_10Y_PUMS:
% python -m sdnist.challenge.submission --test-dataset GA_NC_SC_10Y_PUMS
To score synthetic data files and visualize scores on an interactive map-based html visualizer:
% python -m sdnist.challenge.submission --html
To score synthetic data files during algorithm development (uses public dataset IL_OH_10Y_PUMS):
python -m sdnist.challenge.submission --public --html
The above commands assume that the synthetic data is located in the following directory:
[current-working-directory]/results/census/
.
Each synthetic output file should be named with respect to the epsilon value used for its synthesis.
In its default settings, SDNist performs scoring for epsilons 0.1, 1.0, and 10.0, so the synthetic files would be named
eps=0.1.csv, eps=1.0.csv and eps=10.0.csv, where eps=0.1.csv is synthesized using epsilon value 0.1 and so on.
To generate a final aggregate score over all epsilon values for the Taxi Challenge with a private dataset other than the default:
% python -m sdnist.challenge.submission --challenge taxi --test-dataset taxi2016
The above commands assume that the synthetic data is located in the following directory:
[current-working-directory]/results/taxi/
.
Each synthetic output file should be named with respect to the epsilon value used for its synthesis.
In its default settings, SDNist performs scoring for epsilons 1.0 and 10.0, so the synthetic files would be named
eps=1.0.csv and eps=10.0.csv, where eps=1.0.csv is synthesized using epsilon value 1.0 and so on.
NOTE: The filename of the synthetic data should exactly match the epsilon value provided in the parameters.json file
of the public or private dataset. If an epsilon value mentioned in the parameters.json file is 1
,
then the synthetic data filename should be esp=1.csv
; if an epsilon value mentioned is 1.0, then the synthetic data filename should be eps=1.0.csv
.
The sdnist.challenge.submission
module is used mainly for computing aggregate scores over different
epsilon values, but it can also be used to inspect scores for each epsilon value separately.
To visualize scores over different values of epsilon, year, or PUMA (available only for the Census Challenge):
% python -m sdnist.challenge.submission --html
Other options are available by calling --help
.
2) Option 2 (slightly more advanced and time-consuming)
Reproducing the baselines from the Challenge by subclassing challenge.submission.Model
Some examples of subclassing challenge.submission.Model
are available in the library.
Subsample
Build a synthetic dataset by randomly subsampling 10% of the private dataset:
python -m sdnist.challenge.subsample
Output :
python -m sdnist.challenge.subsample
2021-11-23 14:55:07.889 | INFO | sdnist.challenge.submission:run:66 - Skipping scoring for eps=0.1.
2021-11-23 14:55:07.889 | INFO | sdnist.challenge.submission:run:73 - Resuming scoring from results/census/eps=1.csv.
2021-11-23 14:55:08.007 | INFO | sdnist.challenge.submission:run:88 - Computing scores for eps=1.
100%|███████████████████████████████████████████| 50/50 [00:05<00:00, 9.37it/s]
2021-11-23 14:55:14.969 | SUCCESS | sdnist.challenge.submission:run:92 - eps=1score=842.68
2021-11-23 14:55:14.985 | INFO | sdnist.challenge.submission:run:79 - Generating synthetic data for eps=10.
2021-11-23 14:55:15.565 | INFO | sdnist.challenge.submission:run:85 - (saved to results/census/eps=10.csv)
2021-11-23 14:55:15.565 | INFO | sdnist.challenge.submission:run:88 - Computing scores for eps=10.
100%|███████████████████████████████████████████| 50/50 [00:05<00:00, 9.39it/s]
2021-11-23 14:55:22.530 | SUCCESS | sdnist.challenge.submission:run:92 - eps=1score=842.42
Note that the resulting synthetic dataset is not differentially private.
Random values
Build a synthetic dataset by chosing random valid values:
python -m sdnist.challenge.baseline
This corresponds to the baseline of Sprint 2 (the 2020 Challenge). The output can be considered 0-differentially private if the schema itself is public.
Output:
2021-11-23 14:59:58.975 | INFO | sdnist.challenge.submission:run:79 - Generating synthetic data for eps=0.1.
Generation: 100%|█████████████████████████████████| 20000/20000 [00:32<00:00, 608.57it/s]
2021-11-23 15:00:31.939 | INFO | sdnist.challenge.submission:run:85 - (saved to results/census/eps=0.1.csv)
2021-11-23 15:00:31.939 | INFO | sdnist.challenge.submission:run:88 - Computing scores for eps=0.1.
100%|████████████████████████████████████████████████████| 50/50 [00:05<00:00, 9.64it/s]
2021-11-23 15:00:38.664 | SUCCESS | sdnist.challenge.submission:run:92 - eps=0.1 score=186.73
2021-11-23 15:00:38.682 | INFO | sdnist.challenge.submission:run:79 - Generating synthetic data for eps=1.
Generation: 100%|█████████████████████████████████| 20000/20000 [00:34<00:00, 584.78it/s]
2021-11-23 15:01:12.962 | INFO | sdnist.challenge.submission:run:85 - (saved to results/census/eps=1.csv)
2021-11-23 15:01:12.962 | INFO | sdnist.challenge.submission:run:88 - Computing scores for eps=1.
100%|████████████████████████████████████████████████████████████████| 50/50 [00:05<00:00, 9.50it/s]
2021-11-23 15:01:19.818 | SUCCESS | sdnist.challenge.submission:run:92 - eps=1 score=187.32
2021-11-23 15:01:19.835 | INFO | sdnist.challenge.submission:run:79 - Generating synthetic data for eps=10.
Generation: 100%|█████████████████████████████████████████████| 20000/20000 [00:33<00:00, 596.94it/s]
2021-11-23 15:01:53.417 | INFO | sdnist.challenge.submission:run:85 - (saved to results/census/eps=10.csv)
2021-11-23 15:01:53.417 | INFO | sdnist.challenge.submission:run:88 - Computing scores for eps=10.
100%|████████████████████████████████████████████████████████████████| 50/50 [00:05<00:00, 9.94it/s]
2021-11-23 15:02:00.076 | SUCCESS | sdnist.challenge.submission:run:92 - eps=10 score=186.73
Other examples
Other examples are available in the examples/
folder. DPSyn and Minutemen are directly adapted from the public repos of their authors:
- DPSyn : https://github.com/agl-c/deid2_dpsyn
- Minutemen : https://github.com/ryan112358/nist-synthetic-data-2021. This examples requires the
private-pgm
library (https://github.com/ryan112358/private-pgm)
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.