QueryCheetah: Fast Discovery of Attribute Inference Attacks against Query-Based Systems

Source code repository for the ACM CCS 2024 paper QueryCheetah: Fast Discovery of Attribute Inference Attacks against Query-Based Systems by Bozhidar Stevanoski, Ana-Maria Cretu, and Yves-Alexandre de Montjoye.

We introduce QueryCheetah, the first efficient and scalable tool for automatically finding privacy vulnerabilities in query-based systems (QBS). Our source code allows anyone to thoroughly evaluate the record-specific privacy risk of providing access to a dataset via a QBS. The code is structured as follows:

• main.py is the entry-point running QueryCheetah
• src contains all accompanying utility methods needed for QueryCheetah
• defense contains the implementation of the QBS whose privacy guarantees you are evaluating.

While in the paper we instantiated QueryCheetah against Diffix, the most heavily studied and developed real-world QBS, we cannot release our Diffix implementation here, for intellectual property reasons. We instead release our generic class implementing a QBS together with an easily extendable implementation of a simple QBS, called ExampleQBS, which returns exact query answers (i.e., does not implement any defense).

The method and the code are easily extendable to other QBSs supporting the same syntax: (1) the new QBS needs to be implemented in defense/qbs.py as a class inheriting from the abstract QBS class and (2) the factory method _get_qbs in src/utils.py needs to be updated to return objects of the new class.

1. Python environment

The source code uses commonly used Python libraries. To be able to run the code, first please install the dependencies in the requirements.txt file as follows:

conda create --name querycheetah_env python=3.9
conda activate querycheetah_env
pip install -r requirements.txt

2. Dataset pre-processing

First, run the notebook notebooks/dataset_preprocessing.ipynb. It contains (1) instructions where to find the datasets used in the paper and (2) code to preprocess them once downloaded.

3. Running QueryCheetah

To search for attacks, run main.py by specifying as command-line arguments, such as:

• only_limited_syntax: an indicator if QueryCheetah will search only withing the limited syntax;
• use_mitigations: an indicator if the attacked QBS implements mitigations;
• attack_type: an indicator if QueryCheetah will search for attribute inference or membership inference attacks;
• only_post_hoc: an indicator if the attacks discovered against a QBS without mitigations are applied to a QBS that implements mitigations;
• dataset_name: the name of the dataset;
• qbs: the name of the QBS.

For example, to search for attribute inference attacks in the limited syntax against the ExampleQBS that does not implement mitigations, run:

python3 main.py --only_limited_syntax=1 --use_mitigations=0  --attack_type="aia" --only_post_hoc=0 --dataset_name="adult" --qbs="example"

How to cite

If you re-use QueryCheetah, please cite our paper:

@inproceedings{stevanoski2024querycheetah,
  title={QueryCheetah: Fast Automated Discovery of Attribute Inference Attacks Against Query-Based Systems},
  author={Stevanoski, Bozhidar, Ana-Maria Cretu, and Yves-Alexandre de Montjoye},
  booktitle={Proceedings of the 2024 ACM SIGSAC Conference on Computer and Communications Security},
  year={2024}
}