ADST : Secure Federated Learning

ADST (Asynchronous Distributed Secure Transport) is a privacy-preserving Federated Learning system simulation. It demonstrates how multiple workers can collaboratively train a machine learning model without sharing their raw data or their individual gradient updates with the central server.

🌟 Key Features

• Secure Aggregation: Uses a pairwise masking protocol (similar to Google's Secure Aggregation) to ensure the coordinator can only decrypt the sum of the gradients, not individual ones.
• Differential Privacy (DP): Implements Gradient Clipping and Gaussian Noise injection at the worker level to provide rigorous privacy guarantees.
• Privacy-Preserving:
  • Data Privacy: Raw training data never leaves the worker's local machine.
  • Model Privacy: Individual model updates are masked using cryptographic blinding factors.
• Hybrid Networking:
  • TCP: Used for reliable control messages, key exchange, and peer discovery.
  • UDP: Used for high-throughput transmission of encrypted gradient chunks.
• Real-Time Dashboard: A professional-grade Streamlit dashboard ("Mission Control") to visualize training progress, gradient norms, privacy status, and validation accuracy.
• Cryptography: Built with standard primitives (cryptography library):
  • Ed25519: For digital signatures and identity.
  • X25519: For Diffie-Hellman key exchange.
  • AES-GCM: For authenticated encryption of transport and gradients.

📂 Project Structure

• coordinator.py: The central server. It manages epochs, handles worker registration, and aggregates the masked gradients.
• worker.py: The client. It trains a local CNN on its private data, communicates with peers to generate masks, and sends encrypted updates.
• demo.py: An orchestrator script that automatically launches one coordinator, multiple workers, and the dashboard for a complete local demo.
• dashboard.py: A web interface to monitor the training process.
• data/: Directory containing local datasets for each worker (site1, site2, etc.) and a validation set (val).

🚀 Getting Started

Prerequisites

Ensure you have Python 3.8+ installed.

Installation

Install the required dependencies:

pip install torch torchvision cryptography streamlit pandas matplotlib numpy

Running the Demo

The easiest way to see the system in action is to run the demo.py script. This will start the coordinator, 4 workers, and the dashboard automatically.

1. Start the demo:

   python demo.py

2. View the Dashboard: Open your browser and navigate to: http://localhost:8501

3. Watch it Train:

   • The system will run for a fixed number of epochs (default: 3).
   • You will see workers connecting, performing handshakes, and sending gradients.
   • The dashboard will update with the current epoch, active workers, and training metrics.

4. Stopping: The demo will automatically exit after the training completes. You can also press Ctrl+C in the terminal to stop it early.

🛠️ Configuration

• Number of Workers: You can modify NUM_WORKERS in demo.py.
• Epochs: Adjust MAX_EPOCHS in coordinator.py and worker.py.
• Privacy Budget: Tune DP_NOISE_SCALE and DP_CLIP_NORM in worker.py to balance privacy and accuracy.
• Network Ports: Default ports are 9000 (TCP) and 9001 (UDP). These can be changed in the arguments.

⚠️ Note

This is a demonstration project intended for educational purposes. While it uses real cryptography, it is a simulation running on localhost and is not intended for production deployment without further hardening.