gnnain / fedonce

Practical Vertical Federated Learning with Unsupervised Representation Learning (TBD 2022)

FedOnce

One-Shot Vertical Federated Learning Algorithm

Code Structure

This project contains 4 folders:

• model
• privacy
• torchdp (Cloned from facebook/pytorch-dp, now facebook/opacus)
• utils

Requirements

This project requires conda 4.11.0. The required environment is included in environment.yml. Please run

$ conda env create --name fedonce --file=environments.yml
$ conda activate fedonce

Reproduce the experimental results

Each script is corresponding to a dataset in our experiments

• gisette: gisette
• covtype: covtype
• phishing: phishing
• uj: UJIIndoorLoc
• superconduct: Superconduct
• mnist: MNIST
• kmnist: KMNIST
• fashion_mnist: Fashion-MNIST
• wide: NUS-WIDE
• movielens: MovieLens

Experiment 1: Communication efficiency of FedOnce-L0

Command format: python3 run_<dataset_name>.py For example: python3 run_mnist.py The script will output the accuracy/RMSE for FedOnce-L0 and baselines (Experiment 5 will run after Experiment 1). The communication cost in each round will be recorded under logs/. Then, we use utils/output_loader.py to help load information from logs.

Experiment 2: Performance of FedOnce-L0

Command format: python3 run_<dataset_name>.py For example: python3 run_mnist.py The script will output the accuracy/RMSE for FedOnce-L0 and baselines.

Experiment 3: Performance of FedOnce-L1

Command format: python3 run_dp_<dataset_name>.py For example: python3 run_dp_mnist.py The script will output the accuracy/RMSE for FedOnce-L1 and baselines.

Experiment 4: Performance of FedOnce-L0 on biased datasets

Command format: python3 run_bias_<dataset_name>.py For example: python3 run_bias_phishing.py The script will output the accuracy/RMSE for FedOnce-L1 and baselines.

Experiment 5: Communication efficiency of FedOnce-L0 (more rounds)

Command format: python3 run_<dataset_name>.py For example: python3 run_mnist.py The script will also output the accuracy/RMSE for FedOnce-L0 (more rounds) after finishing Experiment 1. The communication cost in each round will be recorded under logs/. Then, we use utils/output_loader.py to help load information from logs.

Experiment 6: Scalability of FedOnce-L0

Command: python3 run_multi_gisette.py

Experiment 7: Linear-Combine

Command: python3 run_linear.py