Code for the

SECON 2023 paper: VagueGAN: A GAN-Based Data Poisoning Attack Against Federated Learning Systems

A GAN-Based Data Poisoning Attack Against Federated Learning Systems and Its Countermeasure

Data Poisoning Attacks Against Federated Learning Systems

1. Create a virtualenv (Python 3.7)
2. Install dependencies inside of virtualenv (pip install -r requirements.pip)
3. If you are planning on using pca defense, you will need to install matplotlib. This is not required for running experiments, and is not included in the requirements file
4. We retain the interface for label flipping attacks. If you plan to reproduce the label flipping attack, you will need to modify VagueGAN_attack.py

We outline the steps required to execute different experiments below.

1. python generate_data_distribution.py This downloads the datasets, as well as generates a static distribution of the training and test data to provide consistency in experiments.
2. python generate_default_models.py This generates an instance of all of the models used in the paper, and saves them to disk.

• Federated learning system hyperparameters can be set in the federated_learning/arguments.py file
• VagueGAN hyperparameters can be set in the federated_learning/utils/main.py file
• Most specific experiment settings are located in the respective experiment files (see the following sections)

Running VagueGAN attack: python VagueGAN_attack.py

Set in the federated_learning/utils/main.py

Set in the sever.py

Set in the pca_defense.py

Running PCA defense: python pca_defense.py

Running MCD defense: python MCD_detection_metrics.py

If you have any questions or insights regarding this project, please contact us at 21120398@bjtu.edu.cn

If you find this code helpful to your research, please cite our attack or defense paper:

@INPROCEEDINGS{10287523,
  author={Sun, Wei and Gao, Bo and Xiong, Ke and Lu, Yang and Wang, Yuwei},
  booktitle={2023 20th Annual IEEE International Conference on Sensing, Communication, and Networking (SECON)}, 
  title={VagueGAN: A GAN-Based Data Poisoning Attack Against Federated Learning Systems}, 
  year={2023},
  volume={},
  number={},
  pages={321-329},
  keywords={Computer aided instruction;Federated learning;Distance learning;Training data;Generative adversarial networks;Data models;Sensors;Fedrated learning(FL);Security and Privacy;Generative Adversarial Networks(GAN)},
  doi={10.1109/SECON58729.2023.10287523}}

or

@misc{sun2024ganbased,
      title={A GAN-Based Data Poisoning Attack Against Federated Learning Systems and Its Countermeasure}, 
      author={Wei Sun and Bo Gao and Ke Xiong and Yuwei Wang},
      year={2024},
      eprint={2405.11440},
      archivePrefix={arXiv},
      primaryClass={cs.CR}
}