One-Bit Model Aggregation for Differentially Private and Byzantine-Robust Personalized Federated Learning

Implementation Details

This project is implemented using Python 3.10 and PyTorch 1.13. The simulation experiments can be executed through either of the following two approaches:

---

Recommended Approach: Using deepcom Package

We recommend using [deepcom] - a standardized experimental framework specifically designed for communication-efficient deep learning research.

Installation

pip install deepcom  # Install from PyPI

Execution

python -m deepcom config.json

Key Features:

• Automates all experimental configurations
• Ensures reproducibility through parameterized workflows

deepcom Project Repository: mh-lan/deepcom

---

Alternative Approach: Direct Execution with argparse

For custom experimental configurations, directly run the implementation with the following command structure:

python main_fl.py \
 --dataset fmnist \
 --num_channels 1 \
 --model cnn \
 --epochs 300 \
 --num_users 100 \
 --gpu 0 \
 --record_accuracy \
 --repeat 3 \
 --aggregation RSA \
 --attack_type sign_flipping \
 --attack_frac 0.2 \
 --privacy_epsilon 0.1