Federated Learning with Cognitive defence Mechanisms

A modular federated learning framework implementing cognitive defence strategies based on OODA loop and MAPE-K frameworks, with support for various attacks and defences.

Features

• Modular Architecture: Separation of concerns with pluggable attacks and defences
• Multi-Client Orchestration: Automated management of 10+ client processes with resource monitoring
• Multiple Defence Strategies:
  • Cognitive Defence: OODA loop and MAPE-K framework implementation
  • Krum: Byzantine-robust aggregation selecting updates with minimal distance scores
  • Trimmed Mean: Robust aggregation removing outliers from both ends
• Attack Simulation: Label flipping, gradient noise, model replacement, and more
• Explainable AI: Decision logging with reasoning and evidence
• Deterministic Experiments: Reproducible results with proper seeding

Quick Start

1. Setup

# Clone and setup
git clone https://github.com/self1am/FL_CognitiveDefence.git
cd FL_CognitiveDefence
make setup

2. Run Basic Experiment

make run-basic

3. Run Custom Experiment

python -m src.orchestration.experiment_runner --config experiments/configs/your_config.yaml

Defence Strategies

Cognitive Defence

OODA loop-based adaptive defence with reputation system:

• Observes client update patterns
• Orients using historical context
• Decides on weighted aggregation
• Acts with explainable decisions

Krum

Byzantine-robust aggregation from literature (Blanchard et al., NeurIPS 2017):

• Selects client update(s) with smallest distance score
• Robust against up to f Byzantine clients
• Supports both single-Krum and Multi-Krum variants

Trimmed Mean

Byzantine-robust aggregation (Yin et al., ICML 2018):

• Removes β fraction of extreme values per dimension
• Aggregates remaining values
• Parameter-wise robustness

Project Structure

federated-cognitive-defence/
├── src/
│   ├── attacks/          # Attack implementations
│   ├── defences/         # defence strategies
│   ├── clients/          # Client implementations
│   ├── server/           # Server implementations
│   ├── models/           # Neural network models
│   ├── datasets/         # Dataset handlers
│   ├── orchestration/    # Multi-client orchestration
│   └── utils/            # Utilities and configuration
├── experiments/
│   ├── configs/          # Experiment configurations
│   ├── scripts/          # Helper scripts
│   └── results/          # Experiment results
└── tests/                # Unit and integration tests

Configuration

Experiments are configured using YAML files. Example configurations:

Cognitive Defence

experiment:
  experiment_name: "cognitive_defence_test"
  seed: 42
  num_rounds: 10
  server_address: "0.0.0.0:8080"

defence:
  strategy: "cognitive_defence"
  anomaly_threshold: 0.7
  reputation_decay: 0.8

attacks:
  - enabled: true
    attack_type: "label_flip"
    intensity: 0.1
    target_clients: [0, 1, 2]

orchestration:
  num_clients: 10
  batch_size: 3

Krum Defence

defence:
  strategy: "krum"
  num_byzantine: 2      # Expected number of malicious clients
  multi_krum: false     # Use standard Krum (true for Multi-Krum)

Trimmed Mean Defence

defence:
  strategy: "trimmed_mean"
  beta: 0.2            # Trim 20% from each end

Hardware Requirements

• MacBook M1 (8GB): Orchestrator + 2-3 lightweight clients
• Azure VMs (4GB each): Server + 3-4 clients each
• Total: Support for 10+ concurrent clients

Distributed Setup

For multi-machine experiments:

# Setup distributed environment
./scripts/run_distributed.sh

# Monitor progress
tail -f logs/experiment_name.log

Development

# Install development dependencies
make dev-install

# Run tests
make test

# Format code
make format

# Lint code
make lint

Experiment Results

Results are automatically saved to:

• experiments/results/: JSON experiment summaries
• logs/: Detailed execution logs
• Individual client logs with training history

Next Steps

1. FEMNIST Integration: More realistic FL dataset
2. Quantum Neural Networks: PennyLane integration
3. Advanced defences: FreqFed, Median, FoolsGold
4. Adaptive Attacks: Learning-based adversarial strategies
5. Comparative Analysis: Benchmark defences against various attack scenarios

Contributing

1. Create feature branches for new functionality
2. Add tests for new components
3. Update documentation
4. Submit pull requests

License

MIT License - see LICENSE file for details.