Proof of Collaborative Learning: A Multi-winner Federated Learning Consensus Mechanism

Overview

This repository implements PoCL (Proof of Collaborative Learning), a novel blockchain consensus mechanism that replaces energy-intensive mining with federated learning. Instead of solving cryptographic puzzles, miners collaboratively train a global deep learning model, with winners selected based on model performance through a democratic voting system.

Key Innovation

PoCL transforms blockchain mining from a wasteful competition into a productive collaboration where:

• 🧠 Miners train ML models instead of computing meaningless hashes
• 🗳️ Democratic voting determines winners based on model quality
• 🎁 Performance-based rewards incentivize honest participation
• 📊 Global model improvement benefits all participants

🏗️ System Architecture

Core Components

Component	Purpose	Technology
Miners	Train models on local data, participate in consensus	Python + TensorFlow
Blockchain Network	Immutable ledger for transactions and consensus	Hyperledger Fabric
Express Applications	API gateway and process coordination	Node.js + Express
Aggregator	Combine winning models using FedAvg	Python + Flask
Chaincodes	Smart contracts for different system functions	JavaScript

Workflow

graph TD
    A[Transaction Assignment] --> B[Local Model Training]
    B --> C[Model Proposal + Test Data]
    C --> D[Cross-Prediction Phase]
    D --> E[Voting on Performance]
    E --> F[Winner Selection]
    F --> G[Model Aggregation]
    G --> H[Reward Distribution]
    H --> A

Loading

🚀 Quick Start

Prerequisites

• Docker & Docker Compose: For Hyperledger Fabric network
• Node.js 16+: For Express applications
• Python 3.8+: For miners and aggregator
• TensorFlow 2.x: For deep learning models

Installation

1. Install Hyperledger Fabric

   curl -sSL https://bit.ly/2ysbOFE | bash -s

2. Install Python Dependencies

   pip install tensorflow flask requests numpy scikit-learn matplotlib

3. Install Node.js Dependencies

   cd express-application
   npm install

Running the System

1. Start the Complete System

   python3 run.py

   This automatically:

   • Deploys the Hyperledger Fabric network
   • Starts all Express applications
   • Launches 10 miners
   • Initializes the aggregator and submitter
   • Begins federated learning rounds

2. Monitor Progress

   # Real-time monitoring
   tail -f logs/*.txt
   
   # Check specific components
   tail -f logs/app1.txt      # Admin coordination
   tail -f logs/miner1.txt    # Individual miner
   tail -f logs/aggregator.txt # Model aggregation

3. Stop the System

   python3 stop.py

📊 System Parameters

Parameter	Value	Description
Miners	10	Number of federated learning participants
Winners per Round	5	Top performers selected for rewards
Total Rounds	20	Complete federated learning experiment
Training Time	3 minutes	Maximum time for local model training
Prediction Time	15 seconds	Time to predict on others' test data
Voting Time	15 seconds	Time to submit performance votes
Dataset	CIFAR-10	Image classification benchmark

🔄 Consensus Process

Phase 1: Training (180 seconds)

• Miners receive demo transactions to process
• Train global CNN model on local CIFAR-10 data partitions
• Submit trained model hash and test data samples

Phase 2: Prediction (15 seconds)

• Each miner receives test data from all other miners
• Make predictions using their trained model
• Submit predictions to blockchain

Phase 3: Voting (15 seconds)

• Evaluate prediction accuracy on own test data
• Rank other miners based on accuracy and speed
• Submit democratic votes to blockchain

Phase 4: Aggregation

• Select top 5 miners based on vote aggregation
• Combine winning models using FedAvg algorithm
• Distribute rewards proportional to contribution
• Update global model for next round

🏛️ Project Structure

├── 📁 clients/                    # Federated learning participants
│   ├── 📁 miner/                 # 10 individual miners + analysis tools
│   ├── 📁 aggregator/            # FedAvg model combination service  
│   ├── 📁 global model/          # Shared CNN model architecture
│   └── 📁 submitter/             # Transaction generation service
├── 📁 *-coin-transfer/           # Blockchain transaction chaincodes
│   ├── 📁 demo-coin-transfer/    # Demo transactions for mining
│   └── 📁 main-coin-transfer/    # Main cryptocurrency operations
├── 📁 *-propose/                 # Consensus mechanism chaincodes
│   ├── 📁 model-propose/         # Model submission handling
│   ├── 📁 prediction-propose/    # Cross-prediction management
│   └── 📁 vote-assign/           # Democratic voting system
├── 📁 express-application/       # API gateways and coordination
├── 📁 test-network/              # Hyperledger Fabric blockchain
├── 📁 logs/                      # Real-time system monitoring
├── 📁 results/                   # Experimental results and analysis
├── 📁 figures/                   # Architecture diagrams
├── 🐍 run.py                     # Complete system startup
└── 🐍 stop.py                    # Graceful system shutdown

🔒 Security Features

Consensus Security

• Byzantine Tolerance: Resilient to up to 1/3 malicious miners
• Democratic Voting: Equal voting weight prevents centralization
• Model Integrity: Cryptographic hash verification
• Transparent Auditing: All decisions recorded on immutable blockchain

Attack Resistance

• KNN Attack Detection: Identifies miners using simple algorithms instead of deep learning
• Vote Validation: Prevents invalid or duplicate votes
• Deadline Enforcement: Prevents unlimited computation time
• Performance Verification: Cross-validation ensures honest reporting

📈 Performance Results

Training Performance

• Convergence: Models converge within 10-20 epochs per round
• Accuracy: Validation accuracy reaches 70-80% on CIFAR-10
• Efficiency: Complete consensus round in ~4 minutes
• Scalability: Successfully tested with 10 miners, extensible to more

Consensus Quality

• Participation: 90-100% miners participate in each round
• Fairness: Rewards distributed based on actual contribution
• Stability: Consistent winner selection across rounds
• Attack Resilience: 100% detection rate for adversarial miners

Generate Results

cd clients/miner
python3 training_results.py    # Training performance plots
python3 datasize_winners.py    # Data size vs winning analysis

🧪 Experimental Features

Attack Simulation

Test system robustness against adversarial miners:

# In miner1.py and miner6.py, uncomment:
# self.model = KNNClassifier()  # Instead of CNN training

Data Distribution Studies

• Heterogeneous Data: Miners have different amounts of training data
• Two Strategies: Decreasing vs grouped data distribution
• Impact Analysis: Correlation between data size and winning frequency

Consensus Variations

• Winner Count: Adjustable from 1 to 9 miners
• Voting Algorithms: Performance + speed vs accuracy-only
• Aggregation Methods: FedAvg vs other federated learning algorithms

🔧 Configuration

System Scaling

To add more miners:

1. Copy miner1.py to miner11.py (or higher)
2. Add port 8010 to app1.js miners list
3. Update run.py to start the new miner
4. Adjust total_miners parameter in miner configuration

Network Customization

• Modify app1.js: Adjust timing, winner count, round number
• Update miner.py: Change data distribution or model architecture
• Configure aggregator.py: Implement different aggregation algorithms

📚 Research Applications

Academic Research

• Federated Learning: Novel consensus mechanism research
• Blockchain: Energy-efficient mining alternatives
• Machine Learning: Collaborative training in adversarial environments
• Distributed Systems: Byzantine fault tolerance in ML systems

Industry Applications

• Healthcare: Collaborative medical AI without data sharing
• Finance: Fraud detection across institutions
• IoT: Edge device collaboration for smart cities
• Privacy: Machine learning with preserved data locality

🤝 Contributing

We welcome contributions! Please see our Contributing Guidelines for details on:

• Setting up development environment
• Code style and testing requirements
• Submitting pull requests
• Reporting issues and bugs

Development Setup

# Clone repository
git clone https://github.com/your-org/FL-Validated-Learning.git
cd FL-Validated-Learning

# Install development dependencies
pip install -r requirements-dev.txt
npm install --dev

# Run tests
python -m pytest tests/
npm test

📄 Citation

If you use this work in your research, please cite our paper:

@inproceedings{sokhankhosh2024proof,
  title={Proof-of-Collaborative-Learning: A Multi-winner Federated Learning Consensus Algorithm},
  author={Sokhankhosh, Amirreza and Rouhani, Sara},
  booktitle={2024 IEEE International Conference on Blockchain (Blockchain)},
  pages={370--377},
  year={2024},
  organization={IEEE}
}

📞 Support and Contact

• 📖 Documentation: Comprehensive README files in each directory
• 🐛 Issues: GitHub Issues
• 💬 Discussions: GitHub Discussions
• 📧 Contact: [email protected]

📜 License

This project is licensed under the MIT License - see the LICENSE file for details.

⭐ Star this repository if you find it useful for your research or projects!