A Unified Solution for Privacy and Communication Efficiency in Vertical Federated Learning

Published in NeurIPS-2023.

Authors:

Ganyu Wang (Western University)
Bin Gu (Jilin University and MBZUAI)
Qingsong Zhang (Xidian University)
Xiang Li (Western University)
Boyu Wang (Western University)
Charles X. Ling (Western University)

How to run this project

Install the required library

Option 1:
Install pytorch from https://pytorch.org/.

Other packages can be installed via debugging (csv, copy, etc.).

Option 2:
Use the requirements.txt for config the environment. Or use the file as a reference.

$ pip install -r requirements.txt

Run with command line.

$ python main.py --[arguments below] [argument value below]  

Argument	Description	Default	Options
random_seed	Random seed for repeat experiment	12341	-
framework_type	The framework of the VFL (VFL-CZOFO, VAFL, ZOO-VFL)	ZOFO	ZOFO, FO, ZO
dataset_name	dataset	MNIST	MNIST, CIFAR10
model_type	The ML model used for the framework.	MLP	MLP, SimpleResNet18
n_party	number of participants, including the server and clients	3	>1
client_output_size	the dimension of the embedding.	64	MLP: 64, SimpleResNet18: 10.
server_embedding_size	The embedding size of the server's model. Only for MNIST experiment	128
client_lr	The learning rate of the client	0.02
server_lr	The learning rate of the server	0.02
batch_size	Batch size	64
n_epoch	Number of epoch training	100
u_type	The type of the random perturbation of ZOO	Uniform	Uniform, Normal, Coordinate
mu	ZOO parameter	0.001
d	ZOO parameter	1
sample_times	Avg-RandGradEst sampling times (q in the paper)	1
compression_type	The forward message compression type. No compression/Uniform Scale	None	None, Scale
compression_bit	The compression bit of the Uniform Scale Compressor	8	2, 4, 8
response_compression_type	The backward message compression type, No compression/Uniform Scale	None	None, Scale
response_compression_bit	The compression bit of the Uniform Scale Compressor	8	2, 4, 8
log_file_name	The file name of the log file. The log will be saved at the main folder	None

Result

The result you will get is the log output with the training accuracy at each epoch, the communication cost for each epoch, and the test accuracy at the last row.

Citation

If you use or extend our work, please cite the following paper:

@article{wang2024unified,
  title={A Unified Solution for Privacy and Communication Efficiency in Vertical Federated Learning},
  author={Wang, Ganyu and Gu, Bin and Zhang, Qingsong and Li, Xiang and Wang, Boyu and Ling, Charles X},
  journal={Advances in Neural Information Processing Systems},
  volume={36},
  year={2024}
}