Detecting Leaked Data through Synthetic Data Injection and Model Querying

Welcome to this GitHub!

In this repository, we develop a new algorithm, LDSS, to detect leaked data through synthetic data injection and model querying. Shifting local class distribution enables accurate detection of whether a classification model is trained using leaked tabular data.

We devise LDSS with optimization using [FAISS_GPU] (Jeff Johnson, Matthijs Douze, and Herve Jegou. 2021. Billion-Scale Similarity Search with GPUs. IEEE Transactions on Big Data, vol. 7, no. 3, pp. 535–547, 2021). We also include the implementation and comparison to two baseline methods, namely Flip and FlipNN, that perform random label flipping or random nearest neighbor flipping.

Data Sets

Data Sets Details

We study the performance of LDSS on five real-world datasets, i.e., Adult, Vermont, Arizona, Covertype, and GeoNames. Their statistics after data cleaning are summarized as follows.

Datasets	Cardinality	# Categorical Attributes	# Numerical Attributes	# Classes	g	h	Download Link
Adult	48,842	8	5	2	10	44	https://archive.ics.uci.edu/dataset/2/adult
Vermont	129,816	40	8	4	10	118	https://datacatalog.urban.org/dataset/2018-differential-privacy-synthetic-data-challenge-datasets/resource/2478d8a8-1047-451b-ae23
Arizona	203,353	42	8	4	10	185	https://datacatalog.urban.org/dataset/2018-differential-privacy-synthetic-data-challenge-datasets/resource/2478d8a8-1047-451b-ae23
Covertype	581,012	2	10	7	10	528	https://archive.ics.uci.edu/dataset/31/covertype
GeoNames	1,891,513	2	5	9	10	1720	https://www.kaggle.com/datasets/geonames/geonames-database

Data Format

Input data is in CSV format. The first row corresponds to column names. The second row corresponds to column data types: 'C' for categorical features, 'R' for real-valued numerical features, and 'I' for integer features. The remaining rows are data samples with 1 sample per row. The last column is always the class label column with column data type 'C'. Refer to data/adult/adult.data for the Adult dataset and data/arizona/arizona.data for the Arizona dataset as examples.

Requirements

• Ubuntu 20.04 (or higher version)
• Python 3.10 (or higher version)
• GPU with CUDNN available for TensorFlow 2.11 and FAISS_GPU 1.7.2 (or higher version)
• Python libraries listed in requirements.txt

Experiments

We provide the bash scripts to run all experiments. Once you have downloaded the data sets, you can reproduce the experiments by simply running the following commands:

cd models
bash scripts/run.sh

For the respective dataset's experiment script, test\_{dataset}.sh runs LDSS experiments, test\_{dataset}\_b.sh runs Flip and FlipNN experiments, test\_{dataset}\_gh.sh runs parameter study experiments, and test\_{dataset}\_r.sh runs regression experiments.

To run experiments on customized datasets, please run the below command with the right arguments provided.

python model/main.py **args

Some useful arguments are listed below.

Parameter	Description
--dataset	{adult,covertype,arizona,vermont,geonames}, dataset name with respective data file at data/dataset/*.data
--result-dir	result dir
--seed SEED	random number seed
--data-perhole	h, the number of samples to be generated per empty ball
--num-holes	g, the number of empty balls to find and inject data
--test-config	path of testing configuration file
--contamination	contamination for iso forest
--restore RESTORE	whether add random data to restore original class distribution
--split-ratio	k, fold for k-fold experiments
--split-portion	which portion is used as training data (0 to k-1), if split ratio is k-fold
--npivot	m, the number of pivot samples in jacaard distance calculation
--nbit NBIT	k, the number of bits per feature in jacaard distance calculation
--pivot-method	{random,maxfreq,kpp} method of pivot selection
--regression	{T,F} whether to enable regression task, available for Arizona and Vermont for Jaccard method
--eval-multiplier	dilution multiplier to be tested from 1 to k-1 on k-fold (only up to 1 for regression)

Evaluations

Finally, we provide sample Python codes in Jupyter Notebook for evaluations and case studies. Please refer to the folders models/Evaluation.ipynb for more details. To evaluate the experiment results, replace the result file path accordingly.