deki-smpc

deki-smpc is a lightweight client for a Secure Multi-Party Computation (SMPC)-based Federated Learning (FL) framework. Unlike traditional FL frameworks (e.g., Flower) that force you into rigid training loops, deki-smpc lets you seamlessly integrate privacy-preserving aggregation into your own training workflow — whether you're using custom frameworks like nnU-Net or any other deep learning stack.

✨ Key Features

• Minimal Invasion: No need to refactor your training loop! Just plug deki-smpc into your existing code.
• Dual-Server Architecture:
  • Key Exchange Server: Fast, RESTful key generation (built with FastAPI).
  • FL Aggregation Server: Secure model aggregation (leveraging FastAPI (used to be Flower) + PyTorch-based SMPC).
• Efficient Key Generation: Key generation happens in parallel with model training, avoiding extra overhead.
• Security First: Uses a preshared secret and SMPC protocols to guarantee data privacy across participants.
• Flexible and Lightweight: Focused, extensible, and easy to integrate.

📦 System Components

• Key Exchange Server
  REST API using FastAPI for efficient and secure multi-party key generation.

• FL Aggregation Server
  Built on FastAPI with PyTorch-based Secure Aggregation — ensuring privacy without sacrificing flexibility (used to be Flower).

🚀 Quick Usage Example

1. Create a virtual environment:

deki-smpc supports Python 3.10+ and works with Conda, pip, or any other virtual environment. Here’s an example using Conda:

conda create -n deki-smpc python=3.10
conda activate deki-smpc

2. Install this repository

Clone and install this repository:

git clone https://github.com/MIC-DKFZ/deki-smpc
cd deki-smpc
pip install -e .

3. Getting started with the server

First you need to startup a deki-smpc aggregation server.

4. Getting started with the client

from deki_smpc import FedAvgClient

# Initialize the Deki SMPC client
client = FedAvgClient(
    aggregation_server_ip="127.0.0.1",
    aggregation_server_port=8080,
    num_clients=4,
    preshared_secret="my_secure_presHared_secret_123!",
    client_name=client_name,  # For better logging at the server. MUST BE UNIQUE ACROSS ALL CLIENTS
    model=local_model, # PyTorch model
)

# Use it to securely aggregate your model
aggregated_model = client.update_model(local_model)

✅ That's it! Integrate it wherever you train your models.

🔥 Why deki-smpc?

Most FL frameworks dictate your workflow. deki-smpc empowers you to keep your framework (nnU-Net, MONAI, custom PyTorch loops, etc.) while adding secure federated aggregation with minimal changes.

You focus on training great models. We handle secure aggregation.

📚 Documentation

Coming soon! (Stay tuned.)

⚠️ Work in Progress

This is currently a work in progress to transfer the implementation used for the experiments in the paper into a clean and fully usable implementation.

Cite

Hamm, B., Kirchhoff, Y., Rokuss, M., Schader, P., Neher, P., Parampottupadam, S., Floca, R., Maier-Hein, K. (2025). Efficient Privacy-Preserving Medical Cross-Silo Federated Learning. https://doi.org/10.36227/techrxiv.174650601.13181048/v1

Link: