Federated Learning
Training data never leaves local nodes. The federated learning system coordinates model training across distributed participants while defending against malicious updates, enforcing differential privacy, and ensuring data sovereignty. All contributions are compliance-filtered before aggregation.
Federated System
Flower Framework
BelizeChainFederatedClient extends fl.client.NumPyClient for transparent integration with the Flower federated learning ecosystem. Async server architecture handles 100+ participants.
Byzantine Fault Tolerance
ByzantineDetector with 6 defense methods: FedAvg, Krum (single closest update), Multi-Krum (average k closest), Trimmed Mean, Median, and PHOCAS (reputation-weighted aggregation).
Differential Privacy
DP-SGD with ε=1.0 and δ=1e-5. Per-example gradient clipping and Gaussian noise injection ensure provable privacy guarantees for all training participants.
Secure Aggregation
SecureAggregator with masked updates and pairwise key agreement. Noise cancellation ensures the server only sees aggregated results, never individual contributions.
Aggregation Strategies
FederatedAggregator supports FedAvg with three weighting modes (samples, fitness, hybrid) and ByzantineRobustStrategy with trimmed median for adversarial environments.
Validation Nodes
5 initial nodes planned: Government (Belmopan), University of Belize, Karl Heusner Memorial Hospital, Central Bank of Belize, and Toledo Institute for Development and Environment (TIDE).