noobyzy / fedzo

Communication-Efficient Stochastic Zeroth-Order Optimization for Federated Learning

FedZO

---

Black-box aderversarial attack experiments for Communication-Efficient Stochastic Zeroth-Order Optimization for Federated Learning.

For detailed information and results, please refer to the paper.

Requirements

The environment for this experiment is

• Python 3.8.5
• Pytorch

please download other required packages via pip install -r requirements.txt

Dataset

The algorithm is expected to work on multiple datsets, not limited to cifar fashion mnist mnist which are standard datasets

Files

FedZO/
│   README.md
│   requirements.txt    
│
└─── blackbox_attack/
│   │   models/
│   │   save/
│   │   src/
│   
└─── dataset/
    │   cifar/
    │   fmnist/
    │   mnist/

• blackbox_attack/ the main folder -- models/ the well-trained classification models -- save/ the running result of the algorithm, metrics including training loss and testing accuracy -- src/ the algorithm code

• dataset/ the folder stores all datasets -- cifar/ fmnist/ mnist/ folders to store different datasets

src/ contains multiple files:

• alg_*.py the distributed zeroth order optimization algorithm
• attack_main.py the main function to perform aderversarial attack
• load_file.py load dataset and model
• models.py pytorch model structure
• ObjFunc.py blackbox attack objective function
• options.py defining the arguments of the experiments/ hyperparameters of the algorithms
• run_*.sh shell file to excute the experiment (contain default argument setting)
• train_model.py train the DNN classification model from initilization
• utils.py other util functions

Perform the experiment

Once you have configured the environment, run the experiement via

bash run_[your_algorithm_name].sh

and replace [your_algorithm_name] with your desired experiment name.

If everything goes well, you will find the results in save/ folder, each execute of any of the algorithms should yield two products:

• a *.pkl file, produced by pickle package, contains the the algorithm's per-iteration training loss / testing accuracy
• a */ folder sharing the same prefix as the previous file, containing the visualization of the optimized adversarial image, and a mini-batch of samples: the original images and the images perturbed by the adversarial image (aims at showing that the perturbation is imperceptible by human)

Contact Us

If you encounter any problem, feel free to post it in issue of this repository. You can also contact us via the following email:

Ziyi Yu

[email protected]
[email protected]

Wenzhi Fang

[email protected]