Our privacy-preserving federated learning system allows hospitals to collaborate on training a COVID-19 detection model without sharing sensitive patient data. The project is designed with healthcare privacy compliance in mind, following a defense-in-depth security approach with encrypted model weights, secure communication, and proper access controls.

• Privacy-preserving COVID-19 detection using chest X-rays
• Federated learning across multiple simulated hospitals using AWS EC2 instances
• Homomorphic encryption for secure weight sharing using TenSEAL (CKKS scheme)
• Optimized COVID-Net model for encrypted data processing
• Comprehensive evaluation metrics and benchmarking against published works
• HIPAA considerations built into the design
• Scalable architecture that can start small locally and scale to multi-institution deployment

Our system follows a general federated learning architecture with the following key components:

1. Local Hospital Environments: Simulated using AWS EC2 instances
2. Central Aggregator: For secure weight aggregation
3. Homomorphic Encryption Layer: Using TenSEAL for privacy preservation
4. COVID-Net CXR Model: Adapted for encrypted computations

1. Model Initialization: Base COVID detection model created
2. Local Training: Hospitals train on local patient data
3. Secure Aggregation:
   • Local model updates encrypted using homomorphic encryption
   • Encrypted updates sent to aggregation server
   • Server performs addition on encrypted data
   • Result decrypted and applied to global model
4. Model Distribution: Updated global model distributed to hospitals
5. Iteration: Process repeats for multiple rounds

• 1 central server EC2 instance for federated aggregation
• Multiple client EC2 instances representing hospital nodes
• S3 bucket for encrypted model exchange
• KMS for encryption key management

The project follows a staged implementation approach with clear checkpoints:

1. Local Prototype - Simple classification model on medical images
2. Basic Federated Learning - Implementation of federated architecture
3. COVID Detection Model - Scale up to COVID-19 X-ray images
4. Privacy Implementation - Add homomorphic encryption for secure aggregation
5. AWS Deployment - Full system deployment with monitoring

• Homomorphic Encryption: Model weights encrypted with CKKS scheme
• Server-Side Encryption: S3 objects protected with AES-256
• Network Isolation: Security groups limiting communication
• IAM Roles: Principle of least privilege access
• KMS Integration: Centralized key management

The project uses the CKKS (Cheon-Kim-Kim-Song) scheme implemented in TenSEAL:

• Parameters:

  • Polynomial modulus degree: 8192
  • Coefficient modulus bit sizes: [60, 40, 40, 60]
  • Scale: 2^40

• Operations:

  • Encrypted addition of model weights
  • Scale-factor multiplication for averaging

# Clone the repository
git clone https://github.com/ColmCoffey/cxr-secure-federated.git

# Navigate to the project directory
cd cxr-secure-federated

# Install dependencies
pip install -r requirements.txt

# Set up AWS CLI and configure credentials
aws configure

• AWS Account with appropriate permissions
• Python 3.8+
• TensorFlow 2.5+
• TenSEAL 0.3.0+
• Boto3

cxr-secure-federated/
├── client/                     # Client-side code
│   ├── model.py                # COVID detection model
│   ├── training.py             # Local training logic
│   └── encryption.py           # Encryption utilities
├── server/                     # Server-side code
│   ├── aggregator.py           # Secure aggregation logic
│   └── distribution.py         # Model distribution
├── deployment/                 # AWS deployment resources
│   ├── terraform/              # IaC for AWS resources
│   ├── scripts/                # Deployment scripts
│   └── config/                 # Configuration files
├── data/                       # Sample data (not patient data)
│   └── sample_images/          # Example X-ray images
└── docs/                       # Documentation
    ├── architecture.md         # Detailed architecture
    ├── security.md             # Security considerations
    └── performance.md          # Performance benchmarks

Confusion matrix: 
[[194.   6.]
 [ 17. 183.]]
Sens Negative: 0.970, Positive: 0.915
PPV Negative: 0.919, Positive: 0.968
Accuracy = 94.25%

• Encryption Overhead: Benchmarks for homomorphic operations
• Communication Efficiency: Minimizing bandwidth usage
• Model Compression: Techniques to reduce model size
• Scaling Strategies: Horizontal and vertical scaling options

• TensorFlow for model implementation
• TenSEAL for homomorphic encryption
• AWS EC2 for simulated hospital environments
• AWS S3 for secure storage
• AWS KMS for key management
• Python for scripting and data processing

Contact: [email protected]

• Submitting Pull Requests: Follow our code style and include clear descriptions of your changes.
• Reporting Issues: Provide detailed information about the issue and steps to reproduce it.
• Requesting Features: Clearly explain the feature and its benefits.

This project is licensed under the MIT License.

• COVID-Net project for the initial model architecture
• TenSEAL developers for the homomorphic encryption library
• TensorFlow Federated team
• OpenMined for the TenSEAL library

1. Wang, L., Lin, Z. Q., & Wong, A. (2020). COVID-Net: A Deep Convolutional Neural Network Design for Detection of COVID-19 Cases from Chest X-Ray Images. Scientific Reports, 10(1), 19549. https://arxiv.org/abs/2003.09871

2. Wood, A., Najarian, K., & Kahrobaei, D. (2020). Homomorphic Encryption for Machine Learning in Medicine and Bioinformatics. ACM Computing Surveys. https://eprint.iacr.org/2023/1320

3. Brand, M., & Pradel, G. (2023). Practical Privacy-Preserving Machine Learning using Fully Homomorphic Encryption. Cryptology ePrint Archive, Paper 2023/1320. https://eprint.iacr.org/2023/1320

This project was developed as part of a research initiative to improve privacy-preserving machine learning in healthcare settings.