AegisFed: Privacy-Preserving Federated Learning with Oriented Knowledge Purge against Poisoning Attacks

This is the code repository for AegisFed . The "code" folder contains the source code.

📁 Project Structure

code/              	  # Main code directory
├── main.py               # Main program entry
├── server.py             # Server implementation (includes MPC aggregation methods)
├── client.py             # Client implementation (multiple attack types)
├── models.py             # Model definitions
├── data.py               # Data processing module
├── save_results.py       # Results saving module
├── plt_figure.py         # Visualization module 
├── Add_Secret_Share.py   # Secret sharing base module
├── mpc_utils.py          # MPC-KMeans and weight computation utilities

🔧 Environment Requirements

Python Version

• Python 3.7+

Dependencies

pip install torch torchvision
pip install numpy pandas
pip install scikit-learn
pip install matplotlib
pip install openpyxl
pip install backpack-for-python  # For Hessian computation (required by ClientB)

Complete Dependency List

torch>=1.8.0
torchvision>=0.9.0
numpy>=1.19.0
pandas>=1.2.0
scikit-learn>=0.24.0
matplotlib>=3.3.0
openpyxl>=3.0.0
backpack>=0.1.0

📊 Datasets

The project supports the following datasets, which should be stored in the data/ directory:

• MNIST: Handwritten digit recognition (10 classes)
• Fashion-MNIST: Fashion item classification (10 classes)
• CIFAR-10: Natural image classification (10 classes)
• CIFAR-100: Natural image classification (100 classes)

Datasets will be automatically downloaded or loaded from the data/ directory.

🚀 Quick Start

1. Basic Usage

Run federated learning training with default configuration:

cd code
python main.py

2. Command Line Arguments

python main.py \
    --data_name mnist \              # Dataset name
    --num_round 100 \                 # Number of training rounds
    --malicious_ratio 0.2 \          # Malicious client ratio
    --malicious_client_type B \       # Malicious client type (B, C, D, E, F, G)
    --alpha 0.5 \                     # Non-IID degree (None means IID)
    --stop_attack_round 50            # Disable all malicious attacks from this round onwards (optional)

Parameter Description:

• --data_name / -d: Dataset name, options: mnist, fashion_mnist, cifar10, cifar100
• --num_round / -r: Total number of federated training rounds (default: 100)
• --malicious_ratio / -m: Malicious client ratio, range [0,1] (default: 0.2)
• --malicious_client_type / -t: Malicious client type, options: B, C, D, E, F, G (default: B)
• --alpha / -a: Non-IID degree (Dirichlet distribution parameter, None means IID; smaller values indicate more non-uniform data)
• --stop_attack_round / -s: Disable all malicious attacks from this round onwards, clients switch to benign training (default: None, no disabling)

Note: This project is a research project. Please conduct thorough security assessment and performance testing before using in production environments.