This is the source code of paper "Projected Federated Averaging with Heterogeneous Differential Privacy" (accepted by VLDB 2022).
The essential packages for deploying the project:
-
Tensorflow 2.x
pip install tensorflow-gpu
-
Tensorflow Privacy
pip install tensorflow-privacy
or
git clone https://github.com/tensorflow/privacy
- Clone the repo
git clone https://github.com/Emory-AIMS/PFA.git
Note that we omit the basic arguments such as dataset, model, lr, etc. And default Values have been set for these arguments.
- NP-FedAvg algorithm:
python main.py --fedavg True
- FedAvg with HDP algorithm:
python main.py --dpsgd True --eps mixgauss1 --fedavg True
- WeiAvg algorithm experiments
python main.py --dpsgd True --eps mixgauss1 --weiavg True
- PFA algorithm
python main.py --dpsgd True --eps mixgauss1 --proj_wavg True --proj_dims 1 --lanczos_iter 256
- PFA+ algorithm
python main.py --dpsgd True --eps mixgauss1 --proj_wavg True --delay True --proj_dims 1 --lanczos_iter 256
Junxu Liu - [email protected]
Project Link: https://github.com/JunxuLiu/PFA