Awesome

Feature Partition Aggregation

This repository contains the source code for reproducing the AAAI 2024 paper "Provable Robustness Against a Union of $\ell_0$ Attacks".

• Authors: Zayd Hammoudeh and Daniel Lowd
• Link to Paper: Arxiv

Running the Program

To run the program, enter the src directory and call:

python driver.py ConfigFile

where ConfigFile is one of the yaml configuration files in folder src/configs.

First Time Running the Program

The first time each configuration runs, the program automatically downloads any necessary dataset(s). Please note that this process can be time-consuming -- in particular for the weather dataset.

These downloaded files are stored in a folder .data that is in the same directory as driver.py. If the program crashes while running a configuration for the first time, we recommend deleting or moving the .data to allow the program to re-download and reinitialize the source data.

Requirements

Our implementation was tested in Python 3.10.10. For the full requirements, see requirements.txt in the src directory. If a different version of Python is used, some package settings in requirements.txt may need to change.

We recommend running our program in a virtual environment. Once your virtual environment is created and active, run the following in the src directory:

pip install --user --upgrade pip
pip install -r requirements.txt

License

MIT

Citation

@inproceedings{Hammoudeh:2024:FeaturePartition,
    author = {Hammoudeh, Zayd and
              Lowd, Daniel},
    title = {Provable Robustness Against a Union of $\ell_0$ Attacks},
    booktitle = {Proceedings of the 38th {AAAI} Conference on Artificial Intelligence},
    series = {{AAAI}'24},
    year = {2024},
}